AI molecular tumor board

How can we enable timely and cost-efficient access to precision oncology for every patient with complex cancer in the world?

Cancer is one of the most devastating diseases of our time disproportionately affecting lower income as well as ethnic minority populations. An analysis published in JAMA Oncology in February 2023 estimates that, without further investment in research and prevention, cancer will cost the global economy an accumulated $25.2 trillion in health-care costs, lost labor, and spent savings over the next 30 years.

Every cancer is different, so effective treatment strategies can only be enabled by precision (personalized) medicine approaches. Although precision oncology is becoming widely available to the wealthy parts of the population in countries like the US and Switzerland, being technologically and methodologically extremely labor intensive leaves it out of reach for the vast majority of the global population. Aotearoa New Zealand only has the capacity for very few cancer patients to have access to the state of the art precision oncology approaches through a molecular tumor board (MTB), an interdisciplinary panel of experts who advise the lead oncologists on disease state, progression, and treatment options in the individual context.

A typical MTB convenes a group of experts in oncology, pathology, genomics, bioinformatics, pharmacology, and clinical trials. The board is usually presented with a case by a lead oncologist, who is the main point of contact for the patient. A typical case would include information on the clinical history, therapies attempted so far, patient's response, and general therapeutic strategy to date. The board is then asked several questions that the lead oncologist faces. The board’s task is to carry out a comprehensive analysis of the patient history and data and report back to the lead oncologists with advice on their questions. The questions are typically on:

1. Prioritization of approved (targeted) therapies in the context of response (or a lack thereof) to previous treatments

2. Up-to-date clinical guidelines

3. Matching to existing and new clinical trials

The main challenge the board typically faces is the analysis and interpretation of an individual's cancer genomic composition in the context of disease development and progression. This challenge involves deploying algorithms and computational methods capable of associating genes and specific alterations in those genes to response to particular drugs/treatments. Usually these methods are supplemented by literature mining approaches.

With these tasks requiring days of high-cost experts’ time, MTB and precision oncology services are impossible to imagine being accessible to populations without significant financial resources or in remote locations. Even in the US hospitals would typically require the patient to travel to their location before they can make a decision on the patient’s eligibility for their MTB and/or clinical trials.

Despite these significant limitations with respect to populations with MTB access, the global precision oncology market size was valued at over US$100 billion in 2022 and is projected to double by 2030. Making MTBs universally accessible would only intensify this growth.

We aspire to remove the main barrier to precision oncology becoming universally accessible – time human experts spend on linking individuals' clinical and genomic data to biomedical knowledge to come up with clinically actionable recommendations.

Recent developments in AI technology, mainly around natural language processing but also around algorithms for genomics data, promise that the most time-consuming tasks routinely performed by highly qualified human experts can largely be automated to reduce the expert time from days to hours. This can be achieved by a suite of AI algorithms that perform the following tasks:

(1) Process unstructured clinical and molecular information from patient’s records to extract relevant data for bioinformatics algorithms and analyses;

(2) Run relevant bioinformatics analyses on the data;

(3) Link the results of the analyses to clinical and biomedical knowledge, including information on drug trials;

(4) Generate an actionable interactive report for human MTB to interrogate and approve.

We have been working on all these classes of algorithms in the academic framework supported by several NZ Government grants for close to a decade, and we are now at the stage when they can be effectively deployed in clinical practice.

Although most of these tasks are within the reach of modern LLM-based AI algorithms, the main challenge remains (2), that is, developing algorithms that learn clinically relevant information from a patient's molecular data in the context of their individual clinical history. And this is why such algorithms have been the main focus of our research in the past few years. Our algorithms leverage the fact that cancer is an evolutionary process and can be modeled as a game between cancer cells and the environment (initially the immune system followed by rounds of treatment pressure). Snapshots of this game can be observed in the genomic composition of tumors, which in turn can be computationally reconstructed and studied. By reconstructing the evolutionary history of the tumor backwards in time, we are able to identify the history of genomic alternations responsible for various stages of disease progression at the individual level. These genomic alternations then inform drug targets, mechanisms of resistance, and future therapeutic strategies, which combined form the basis of MTB reports. Combined with modern LLMs and our own algorithms capable of handling new data on-the-fly (known as online algorithms), we are currently extending this system to work in the interactive framework for human experts to be able to request additional information, challenge model assumptions, probe different therapy scenarios, etc.

Primarily our solution will serve cancer patients that are currently excluded from the world's most advanced healthcare systems. The three main groups we're developing our technology for are populations with limited access to cancer healthcare due to (1) cost; (2) location; (3) ethnicity. In the Aotearoa NZ context we are working with Māori (indigenous people of Aotearoa) and Pasifika populations, who are currently to large extent excluded from precision medicine due to genomics and algorithms developed for and trained on people of European descent. We are developing algorithms trained on large genomic datasets (e.g. UKBB) that we then finetune on the target population genomes. With precision medicine shifting focus from population-level approaches to individual level, ethnic minorities should be at the forefront of populations who benefit from this technology.

Cancer is a disease where delays in decisions/treatment can mean the difference between life and death. Because our technology would shorten the time MTBs spend on an individual case, every patient with complex cancer will benefit from this.

By region, North America has generated more than 43% of revenue share of the global precision oncology market in 2022, with Asia-Pacific expected to expand at the fastest CAGR between 2023 and 2032. So these two regions are the obvious markets to benefit from our technology.

By cancer type, every third cancer in women is breast cancer. With surgery being a very common approach to breast cancer treatment, patients with this condition naturally form one of our main target populations because molecular sequencing data can be extracted from both biopsy samples and removed tumors and metastases. Time-resolved samples from multiple locations are the perfect type of data for our technology. Breast cancer is of course not the only type of cancer where this is true — colorectal cancer, the third most common cancer worldwide, is another such type.

We are a group of computer scientists, molecular biologists, and oncologists. With decades of experience among us, we are internationally recognized experts in our respective fields, driven by the common goal of improving cancer care accessibility for everyone.

The four crucial areas that need to be covered for this endeavor to succeed are:

1. Biomedical data science

2. Algorithms

3. Oncology

4. Indigenous health

Our team includes an internationally recognized expert in each of these areas. It took us years to build this widely transdisciplinary collaboration, with very significant effort going into enabling effective communication between all of us. We are now a team where everyone understands enough of everyone else’s discipline to be able to communicate effectively and efficiently.

We interact first-hand on a nearly daily basis with populations that will benefit from our technology — cancer patients with complicated clinical situations including indigenous and other people of non-European ancestry, their oncologists, pathologists, and the New Zealand Molecular Tumor Board. We hence are favorably positioned to deliver the impact we are aiming to achieve.

Aotearoa New Zealand has robust processes of engaging with indigenous communities in culturally sensitive ways. Across our team we have years of experience working with Māori and Pasifika communities to guide our work by their input, ideas, and interests.

Large-scale multimodal AI systems are taking humanity by storm. Although their capabilities are vastly promising and limitations yet to be understood, the utility of these systems in clinical oncology requires several inflecting points of innovation to succeed. We are yet to see systems of algorithms performing tasks as complex as those performed by molecular tumor boards. Building on existing AI technologies to enable highly accurate and reliable predictions in the clinical context lies at the heart of our innovation. We will catalyze the effectiveness of AI systems built for a relatively narrow task but having zero tolerance for errors. The heterogeneity and unstructuredness of clinical data demand highly flexible multimodal algorithms capable of retrieving information from a patient's genome, clinical records, pathology images, etc., matching this information with relevant biomedical knowledge, summarizing and interpreting relevant knowledge back into the individual patient’s context. All recommended decisions need to be backed by references, clinical and drug trial guidelines.

Leveraging the fact that cancer can be seen as an evolutionary process and applying both evolutionary and game-theoretic approaches to identifying disease drivers and drug targets is the bleeding edge of research in computational oncology, to which our own research has already contributed. This project will translate this research into the clinic.

The global precision oncology market size was valued at US$100.06 billion in 2022 and is projected to double by 2030. Our AI MTB with its universal access would intensify this growth. With AI technologies penetrating all aspects of our life and creating significant value in the relevant markets, we expect the same to be true for oncology — with AI technologies making various aspects of oncology healthcare more efficient additional market value will be created as well as existing market composition will change.

The therapeutics segment has contributed the largest revenue share of 72% in 2022 to the global precision oncology market. AI is shifting the drug discovery paradigm, with all large pharmaceutical companies committing significant investments in AI. Our technology would naturally tap into this market by allowing identification of drug trials with the highest chance of benefitting an individual patient. While saving lives, the technology can as well support drug trial campaigns.

Quite literally — our solution will ensure healthier lives for those who otherwise would be excluded from quality healthcare, independently from the place where they live, their socioeconomic status, ethnic background, or age.

Our AI system has three components: algorithms that learn from molecular data to identify cancer driver mutations and drug targets at the individual level, algorithms that link a patient's clinical history to biomedical knowledge, and a multimodal learning approach that combines all these sources of information to leverage interactions between them. The “What is your solution?” section has more details on how these systems work.

We are using large-scale cancer datasets (UKBB and others), as well as our own data from the Aotearoa New Zealand region. We are also establishing partnerships with hospitals in Australia and NZ to access their molecular and pathology data. We envision developing a model, where we use the data from our patients for training future generations of our algorithms.

We are operating in a very regulated environment, and we ensure all ethical, responsibility, and privacy regulations are strictly followed. These include patient privacy, data and information handling, and indigenous data science practices among others.

Our systems are not yet at the level of imposing any risk that would be serious enough to actively mitigate. The main risk currently is a patient data breach. We mitigate this risk by enforcing the industry standard cyber security systems to protect our patients' data. Our AI system has the potential of getting close to general biomedical intelligence, which will come with widely ranging risks. We closely monitor capabilities of our system and will develop a mitigation plan when those capabilities will have a chance of emerging in a realistic time frame.

Impact Goal 1 (2024): Ensure that AI predictions are clinically actionable.

We will achieve this goal by developing collaborations with a diverse network of oncologists and their patients. We have done this work in Aotearoa New Zealand and Australia and have a good understanding of the molecular tests and reports requested by oncologists, and their need for data analysis / interpretation of those tests. Within this program we will investigate the unique healthcare system of the US, identify our target audience that can benefit from AI MTB support, and perform multiple case studies. MTBs are more common in the US than they are in Australia and NZ, so we’re planning to connect to North American MTBs to learn about their practices and opportunities for our AI to make those practices more efficient.

Impact Goal 2 (2024): Indigenous genomics.

We will continue our work aimed at identifying genetic drivers of cancer in Māori and Pacifika populations through fine-tuning deep learning models trained on large datasets such as UKBB. The Cure residency would be invaluable for us to connect to the East Coast researchers working in this area of minority populations genomics, as well as to reconnect to our exciting collaborators in this area in Seattle, WA. We hope to establish partnerships with those researchers around data sharing and algorithm development expertise.

Impact Goal 3 (2024-2028): MTB services globally.

Healthcare is a highly regulated industry, with different countries employing significantly different approaches to healthcare regulation. As we aspire to make our service available to every cancer patient in the world, we are developing an approach that will cater to a variety of healthcare systems. Our current model integrates NZ and Australian systems, both of which rely significantly on publicly funded (through general tax) healthcare. The Cure residency will enable us to develop an extended model that is appropriate, beneficial, and economically viable for the US healthcare system. Specifically, we hope that Cure mentors can support us developing a strategy for finding product-market fit in the North American market with the ultimate goal of becoming a global technology.

Five people are currently working on our solution, one of whom is full-time.

We are professors with 10+ years experience in research directly relevant to our solution. We started actively developing Allagi.ai in 2022.

We have a robust process to ensure that every new member of our team is selected based on their abilities and independently of their social background, ethnicity, gender, sexual orientation, etc. We only assess achievements weighted by the opportunity, and this approach has proven very successful. As a result, we are a very productive team representing multiple ethnic backgrounds, sexual orientations, and 60% women.

Allagi.ai is an early-stage techbio startup in stealth mode. We are five academics working towards completing the MVP in Q3 2024 and establishing the company independently from our universities. At that stage at least two of us will be full-time with Allagi.ai and we’ll start fundraising and scaling operations.

We are leveraging our network to engage with our main stakeholders and customers — hospitals, pathology labs, and patients. Through this engagement we learn about the market demand and iterate our product to meet our customers specific needs. We also work with NZ Cancer Trials to understand how we can partner with biotech companies recruiting patients for their trails. Since NZ drug trial market is very small, this is something we’re planning to do a lot more of when we expand to bigger markets such as the US.

The main resources required to successfully build and implement our solution are (1) AI, genomics, and oncology expertise, (2) computational resources, and (3) business/startup development expertise. We are well-resourced in (1) and (2) at the moment as we have all necessary resources originating from our labs. As we grow, we’re planning to tap into larger job markets (North America, Asia, and Europe) for (1), and integrate our local servers with the cloud via Azure to address (2). Participating in the Cure Residency would be an invaluable contribution to developing our capabilities in (3).

So far we have been successful in funding our work from government grants. In the future we are planning to continue applying for grants and seeking donations to fund our R&D activities. As these are limited and do not scale to developing our solution globally and delivering the impact we are aiming for, we plan to raise investment capital to fund the commercialization work. We envision our commercial product to be a tiered service of tracking an individual's cancer risk/status over time. Specifically, we link an individual’s genomics and clinical history to today’s biomedical knowledge, run our algorithms and analyses, and deliver this knowledge to the client depending on their tier:

1. Detailed report containing all biomedical information relevant to the individual in precise scientific language. Typical customer: research hospitals, pathology labs, molecular tumor boards, biotech companies

2. Clinically relevant and actionable information and visualizations for non-academic oncologists. Typical customer: non-research hospitals/practices and pathology labs

3. Basic information and visualizations similar to https://www.cancer.gov/. Typical customer: individuals keen to track biomedical progress in their personal context. These individuals can request tier 1 or 2 services delivered to their oncologist

With these offerings we will tap into the global precision oncology market, which was valued at over US$100 billion in 2022. The breast cancer segment, one of our main focuses, has generated more than 42% of the revenue share in 2022. By product, the therapeutics segment has contributed the largest global revenue share of 72% in 2022. Our technology would naturally tap into this segment of the market by allowing identification of drug trials with the highest chance of benefitting an individual patient. So we envision partnering with biotech companies carrying out cancer drug trials. Our algorithms will be adding value to those companies by identifying patients who are likely to benefit the most from a particular trial.

We are bootstrapping the company so the majority of our operating costs come from team members donating their time and personal computational resources to work on this project. We have projected our annual burn rate to be $200k in 2023; $400k in 2024 (two full-time employees); $1m in 2025 (following the seed round). In addition to the founding team, we expect to employ an AI engineer and a biomedical data scientist following the seed round.

To continue our work in 2024 we are seeking funding of $100k.

Techbio spin-outs from Aotearoa NZ universities is a recent phenomenon, and being one of very few of these companies in the period of high interest rates when early stage biotech venture capital is effectively evaporating is challenging. Hence, we have been operating on an extremely lean budget and every dollar spent at Allagi.ai goes a long way.

This award  would allow us to commit two members of the founding team to work on this project full time, develop our MVP, and launch our seed fundraiser in Q3 2024. We will use this funding to establish partnerships with hospitals, pathology labs, and MTBs in the NYC area. We will achieve this by iterating on our product to get it to the stage when our major partners can make LoI commitments. To make this possible, we will relocate at least one of our team members to NYC for the duration of the Residency.

Cure Residency would be a perfect opportunity for us to gain hands-on experience with the North American healthcare system, engage with hospitals, pathology labs, molecular tumor boards, and patients in the region, and receive invaluable mentoring from Cure's Executive Advisory Board and networking support crucial for an early-stage company. NYC is one of the world’s main centers for both AI and biomedical research with a thriving startup ecosystem. For a company originating from Aotearoa NZ becoming part of this ecosystem would provide the exact kind of acceleration we are currently looking for, so we are particularly excited about the networking and mentorship opportunities.