AI Tool CHIEF Paints a Landscape of a Cancer, Refining Diagnosis, Treatment, and Prognosis

A diagnosis of stage IV cancer used to mark the beginning of the end. Today for many patients, it is the beginning of taking a series of drugs that specifically target the errant cells by blocking the signals that fuel their runaway cell division, while sparing healthy cells. Stage IV cancer patients can live years, even decades, sometimes succumbing to something else.

Now there’s hope even for patients whose cancers become resistant to targeted drugs – using artificial intelligence to probe cancer cells and their surroundings to identify novel points of vulnerability. Researchers from Harvard Medical School describe a new ChatGPT-like model that can guide clinical decision-making to diagnose, treat, and predict survival for several types of cancer. Their report appears in Nature.

The new approach complements targeted drugs by going beyond a cancer cell’s surface and biochemical pathways within, to also probe the microenvironment – the immediate surroundings – through image analysis. If deployed early, AI might identify drugs unlikely to work more effectively than can genetic and genomic testing. It is a “can’t see the forest for the trees” strategy, revealing the landscape of a cancer.

A Brief History of Targeted Cancer Drugs

In 1978, FDA approved the first targeted cancer drug, tamoxifen. Rather than killing all rapidly-dividing cells like traditional chemo, tamoxifen blocks estrogen receptors. And that keeps out the hormone that triggers too-frequent division of breast cells.

The FDA approved Herceptin, targeting a different receptor (HER2) in breast cancer cells, in 1998.

The most groundbreaking cancer drug approval was for Gleevec, with a mere 3-month trajectory at FDA before its approval in 2001. Gleevec is a small molecule that interferes with entry of an enzyme – a tyrosine kinase – that enables growth signals to enter specific cells and trigger division. It was the first kinase inhibitor, initially used to treat a form of leukemia.

Targeted cancer drugs developed to treat one type of cancer often expand, with further testing, to treat other types, based on molecular similarities. Diagnosis began to shift from a body-part basis to a molecular one, which I wrote about in “Mutation and location important in cancer treatment” for The Lancet in 2015.

And so Gleevec’s reach rapidly expanded to treat other blood disorders and a digestive cancer, which saved the life of a friend of mine.

Keytruda illustrates more recently approved cancer drugs. FDA-approved in 2014, Keytruda was the first “programmed death receptor-1 (PD-1) inhibitor.” Use expanded from melanoma to kidney, lung, uterine, and head and neck cancers – but specific molecular subtypes of these cancers.

Translating the Runaway Jargon in Cancer Drug Ads

Ads for cancer drugs come with staccato gibberish meant to describe molecular mechanisms and biochemical pathways, antibodies and their receptors, as if the average distraught patient can instantly envision the steps of a signal transduction pathway or an antibody binding an antigen.

To help, the final syllable of an unpronounceable drug name indicates drug type.

A name ending in “nib” means a small molecule that inhibits an enzyme called a kinase, and is short for “inhibit.” Because the enzyme is required for growth signals to enter and trigger cell division, the drug stops the runaway dividing. Gleevec, aka imatinib, does this.

A name ending in “mab” is a monoclonal antibody, abbreviated MAb, which acts like a drone. Keytruda (pembrolizumab) binds either of two molecules (programmed death receptor-1 [PD⁠-⁠1] or the programmed death ligand 1 [PD⁠-⁠L1]). This removes the cancer’s blocking of the immune response, so that T cells can fight the cancer. Keytruda’s classification based on mechanism, rather than its structure, is immune checkpoint inhibitor.

AI Probes the Cancer Microenvironment

Targeted cancer drugs are based on genotype – the sets of mutations that initiate and drive the cancer. The new AI approach does that, but also considers phenotype – what the tumor and its environs look like, analyzing digital slides of tumor tissues. It’s called CHIEF, for Clinical Histopathology Imaging Evaluation Foundation.

The researchers tested the model on 19 types of cancers, and validated the findings among several international patient groups.

“Our ambition was to create a nimble, versatile ChatGPT-like AI platform that can perform a broad range of cancer evaluation tasks. Our model turned out to be very useful across multiple tasks related to cancer detection, prognosis, and treatment response across multiple cancers,” said senior author Kun-Hsing Yu.

The AI model reads and analyzes digital slides of tumor tissues, identifying cancer cells and predicting a tumor’s molecular profile based on cellular features and characteristics of the surrounding area, aka the microenvironment. These findings can then be used to predict response to treatments as well as survival time.

Perhaps most importantly, the team said, the AI tool can generate novel insights, such as sets of tumor characteristics previously unrecognized to predict patient survival.

“If validated further and deployed widely, our approach, and approaches similar to ours, could identify early on cancer patients who may benefit from experimental treatments targeting certain molecular variations, a capability that is not uniformly available across the world,” Yu said.

A Huge Data Dump

AI synthesizes and links immense volumes of data, and makes projections.

The researchers trained CHIEF on 15 million unlabeled images grouped by tissue type or location in a particular organ or structure. Next, CHIEF trained on 60,000 more images, representing many body parts: lung, breast, prostate, liver, brain, colon, stomach, esophagus, kidney, bladder, thyroid, pancreas, uterus, testes, skin, adrenal glands, and more. The training considered location context – that is, exactly where a particular cell lies within the 3D space of a tissue or organ.

The training strategy powered CHIEF to interpret an image in a broad context as well as focusing on a specific part of an organ.

After training, CHIEF was given more than 19,400 whole-slide images from 32 independent datasets, collected from 24 hospitals and patient cohorts from all over the planet.

Findings

CHIEF performed well no matter how a sample was obtained (surgery or biopsy) or the image digitized, indicating feasibility in diverse medical settings. The tool detects cancer cells and pinpoints the origin of the tumor, as well as identifying and assessing gene and DNA patterns to predict patient outcomes and treatment responses. It even worked on slides that hadn’t previously been analyzed or cancer type identified.

Analyzing DNA patterns using AI may replace time-consuming process DNA and genomic sequencing tests, Yu said. Instead, CHIEF deduces genotype – mutations –by considering association with the phenotype – the images.

CHIEF also enables a clinician to immediately evaluate targeted drugs for a particular patient. The team tested this for predicting mutations in 18 genes, across 15 body parts, linked with response to FDA-approved targeted therapies.

Perhaps most importantly, CHIEF scored well in predicting patient survival, from images of the tumor sampled during the initial diagnosis. It can distinguish tumors surrounded by immune cells (indicating long-term survivors) from tumors devoid of these protective cells (shorter-term survivors).

CHIEF also visualizes a particular tumor’s aggressiveness, using heat maps, or “AI-derived hot spots,” that indicate cancer cells interacting with neighboring non-cancer cells – a sign of imminent spread.

Tumors from patients with a poorer prognosis also tended to include cells of different sizes, telltale lobed cell nuclei, weak links between cells, dying cells, and less dense connective tissue networks, CHIEF found. Cancer clearly entails more than just errant genes.

The researchers plan to expand their analysis of CHIEF to:
• analyze non-cancerous conditions, especially rare diseases that are challenging to diagnose and treat
• detect pre-cancerous cells
• predict levels of aggression of cancer cells at different stages
• identify benefits and adverse effects of novel cancer treatments

As someone who has had thyroid and breast cancer, I applaud the wedding of AI to DNA-based cancer diagnostics. More information enables more treatment choices.