AI Meets Molecular Biology, Genetic Oncology and Pharmacogenomics

Recent advances in AI have led to unprecedented breakthroughs in science. In structural biology, models like DeepMind’s AlphaFold solved the decades old problem of protein structure prediction, paving the way for advances in the molecular structure-function paradigm and drug discovery. In chemistry, AI has enabled the design of novel molecules with specific properties, exemplified by generative models producing innovative drug candidates and materials.

Building on these achievements which underscore AI’s growing influence in scientific discovery and advancement, Periculum is proud to announce our interdisciplinary research across molecular biology, genetic oncology, pharmacogenomics and AI. We are carrying out studies at the molecular and multiomics levels to discover how deep learning and generative AI can enable more precise genomic profiling, characterize mutation effects, and uncover targeted therapies for various cancer types. This convergence is not only enhancing our understanding of complex biological systems but will also enable us to introduce scalable, personalized AI solutions to global health challenges, especially cancer.

Our Motivation

Breast cancer is currently the deadliest cancer type in Nigeria, followed by cervical cancer, with estimated incidence of 25.7% and 14.6%, respectively, and estimated mortality proportions of 18.6% and 16.8%, respectively. These cancer types present some of the most significant regional and global health challenges, with early detection, accurate characterization, and effective treatment being critical to improving patient outcomes. 

Some of the central issues that hinder cancer detection, prevention and care include: 

• Silent cancer development and non-specific symptoms for early-stage cancers pose a unique challenge in early detection and treatment. Vague symptoms such as fatigue or mild pain can be mistaken for benign conditions. As a result, patients often seek medical attention only when the disease has progressed.
• Cancers are highly heterogeneous. The same tumour can characterize different genetic, epigenetic and molecular profiles which makes it difficult to accurately characterize tumors and predict their behavior.
• Inadequate screening, lack of reliable screening tools, and late diagnosis, with many cases diagnosed at advanced stages, potentially increasing the mortality rate.
• Limitation of current diagnostic tools, such as biopsies and imaging which can miss small or less accessible tumors, and their accuracy may be limited by the skill of the practitioner.
• Underutilization of genetic data for molecular profiling and analysis of patients' DNA signatures which can enhance early detection and personalized treatment.
• Challenges in effective and targeted treatment, and limited access to precision medicine which involves tailoring treatments based on the genetic and molecular profile of a patient's tumor.
• Cancer cells can develop resistance to chemotherapy, targeted therapies and immunotherapies. Resistance often arises due to genetic mutations that allow cancer cells to evade treatment.

These challenges present unprecedented opportunities for integrating AI to enhance cancer care through precise analysis and interpretation of genetic, epigenetic, and molecular information. By bridging molecular and genomic research with AI, we are designing AI algorithms for analyzing the genetic mutations that cause breast and cervical cancers, uncovering how these mutations alter the structure and function of their associated proteins and drive oncogenesis (i.e., cancer generation). These algorithms will be capable of predicting the likelihood of oncogenesis and unlocking preventive therapies for breast and cervical cancers before they form. In other words, we are advancing pre-cancer detection and preventive care with AI!

Research and Development Objectives

Our interdisciplinary work drives convergence across AI, molecular biology, precision oncology and pharmacogenomics. Through this large-scale project, our aim is to build intelligent applications capable of performing the following:

• Detect whether proteins encoded by breast and cervical cancer genes (BRCA 1, BRCA 2, TPR53, etc) have been mutated by analyzing their amino acid sequence.
• Predict the effects of gene mutation on structural and functional changes of the associated proteins compared to their canonical (wild-type) versions. This includes detecting changes in protein subcellular localization and activity.
• Through proteomic and mutation analyses, predict the likelihood of breast and cervical cancer generation and, potentially, how long it will take before oncogenesis. 
• Uncover clinically-tested and FDA-approved drug candidates for inhibiting or slowing down oncogenesis if applicable, or provide suitable treatment plans otherwise, based on genomic signatures of the patient.
• Design new experimental molecules for (i) cancer prophylaxis, (ii) reversing mutation effects and oncogenic progression, and (iii) enzymatic targeting in which isomerase enzymes could target mutated proteins to create conformational changes by reversing their folded structure (See here and here, for example).

Multi-Phased AI Experiments

As part of our research, we are designing and running multi-phased AI experiments with a wide variety of algorithms and architectures, to enable us to implement capable AI systems that empower clinical oncologists, medical lab scientists, and pharmacology researchers to utilize AI-powered precision oncology for clinical and research applications. These experiments include training and finetuning of:

• Open-source large language models (in particular, Llama 3+ family of models) on large-scale biomolecular datasets for multiomics analysis.
• Deep learning architectures for protein structural alignment, secondary structure prediction, and membrane association prediction.
• Genomic and proteomic embedding models.
• Molecular docking and drug-binding algorithms.

LLMs are crucial for our research not only because of their ability to unlock valuable insights from data but most importantly because they can understand the language of life encoded in DNAs. On one hand, we are finetuning Llama for protein structure-function-interaction analyses, gene mutation analysis and interpretation, and candidate drug prediction. On the other hand, we are implementing retrieval augmented generation (RAG) pipelines with the pretrained Llama models to facilitate contextual analysis with clinical data and external knowledge bases, enabling grounded interpretation of predictions and contextual conversational interactions.

People and Partnerships

This interdisciplinary effort, led by Zion Pibowei, our Head of Data Science and AI, brings together a cross-functional team of bioinformatics researchers, data scientists, machine learning engineers, and software engineers. Periculum is also proud to collaborate with NSIA-LUTH Cancer Centre (NLCC), the largest cancer center in West Africa. NLCC has a stellar reputation for groundbreaking work in oncology, advanced research capabilities, and a robust patient database that provides invaluable insights into cancer diagnosis, treatments and outcomes. 

As the first center in West Africa to integrate AI into contouring, treatment planning and patient monitoring, NLCC has significantly improved both the efficiency and effectiveness of radiation therapy for its patients. This partnership allows us to leverage NLCC's expertise and resources to develop AI-powered precision oncology using patient-specific clinical data.

We will also continue to foster stakeholder participation and buy-in across the research and development lifecycle with key players in the cancer healthcare regulation and management ecosystem, such as NICRAT and NCCP.

Through these partnerships, we aim to revolutionize cancer care in West Africa, bringing cutting-edge personalized therapeutics to patients and improving healthcare outcomes at scale.

‍