AI Everywhere

Translating the power of AI into practical and usable solutions within and beyond R&D in the life sciences domain

A lot changed in 2023. The widespread use of AI through ChatGPT changed perceptions of the use and relevance of AI in creative tasks. For example, we are already harnessing AI to write code to create end-to-end pipelines and applications more flexibly and easily. We are already using AI in creating user-centric applications, with natural language interfaces.

Beyond its ability to seemingly understand questions and generate responses in natural language, we at Aganitha see an opportunity to make AI understand and speak the language of life – DNA, RNA, protein sequences, molecules, genes, targets, drugs, and so on.

Generative AI-based platforms for therapeutic design

Firstly, we are using generative AI and foundational models for therapeutic design to develop multiple in silico discovery platforms.

1. Platform for de novo generation of antibodies: Our generative AI models jointly model sequence and structure, generating target epitope-specific and, equally importantly, developable antibodies with the right physico-chemical properties needed in therapeutic mAbs. Combined with deep structural modeling and molecular dynamics capabilities, we are able to tackle difficult and novel targets for which structure and activity mechanisms have not yet been fully solved.
2. Platform for de novo generation of small molecules and PROTACs: Our generative models filter the vast ‘drug-like’ chemical space to optimize for biological activity, synthesizability, drug-like characteristics, and ADMET properties. These models mimic the fragment-based drug discovery approaches in the lab that use ultra-large combinatorial synthesis libraries (CSL). We extended our methods to also design protein-degraders (PROTACs).
3. Platform for Viral vector design in Cell and Gene therapies: Our AI models learn from and extrapolate the results of large, multiplexed directed evolution experiments to optimize for tropism to target cells and reduce off-target effects. Our models scrutinize samples for vector-host integration and evaluate the risk of oncogenicity, a common requirement now for all approved gene therapies.

AI models for synthesis and process management

Secondly, we are developing AI models beyond discovery use cases, addressing synthesis and process management concerns of CMC teams:

1. Models for Biosynthesis reactions: Our AI models are forecasting the titer of antibody synthesis reactions accurately and early, enabling better operational planning in downstream facilities. Our AI models can also reliably predict the physico-chemical properties of synthesized enzymes and proteins, factoring in the impact of synthesis conditions.
2. Models for cross-coupling reaction catalysis: We take advantage of GPU implementations of DFT-powered Quantum Chemistry simulations to accelerate the study of cross-coupling reactions catalyzed by transition metal complexes. Optimizing yield and avoiding undesirable side products are the key goals served.
3. Models for troubleshooting gene therapy synthesis: The FDA has recently identified solving manufacturing issues in advanced cell and gene therapies as a very important priority over the next decade. Our work with long-read sequencing is enabling better investigation of issues such as empty capsids, vector truncations, transgene, and UTR defects.

LLMs and foundation models for user experience

Thirdly, we are able to bring the fruits of LLMs and foundation models in Biology to improve the scientist experience and reduce cycle times in research. Specifically:

1. Interrogate your Omics dataset: Our LLM prompt libraries provide conversational interfaces to complex genomics, transcriptomic, and multi-omics datasets, enabling scientists to interrogate and analyze the results of Omics analyses such as GWAS, scRNA-seq, and perturb-seq beyond what traditional dashboards with tables and visualizations enabled.
2. Leverage AI foundation models of biology: What gene is a non-coding variant impacting? And what is the direction of impact? How about deciding on the right target for ASO therapies addressing splicing disorders, factoring in the need for testing in model organisms as well? Do you want to simulate the knockout of a candidate target gene and understand the impact on other genes and pathways? We are able to accelerate these inquiries using AI foundation models for DNA, RNA, and Proteins.
3. Target analysis and market intelligence: Researchers are challenged with keeping track of and evolving insights from the ever-increasing data – research papers, datasets, and even industry news. Our pipelines use fine-tuned LLM models and tested LLM prompt libraries to help the researchers get specific insights relevant to target analysis, disease studies, and even market intelligence.

Clearly, AI is everywhere. To fully realize its potential, we are transforming our training, solution development, and DevOps in the company. Our tools should help us use generative AI within your own premises, with your own data, to provide exploratory interfaces for R&D and beyond. To know more, click here.