Regina Barzilay: Deep Learning for Cancer Diagnosis and Treatment

At a Glance

Regina Barzilay is a professor at MIT and a leading researcher in the application of natural language processing and deep learning to chemistry and oncology. The main theme of this episode: the potential of deep learning in early cancer diagnosis and drug design, as well as the barriers to technological deployment posed by data acquisition and regulatory frameworks.

The most impactful judgment in the entire episode: Barzilay believes that current AI algorithms for cancer diagnosis are already "orders of magnitude" better than clinical standards (such as breast density assessment), but the real bottleneck lies not in the algorithms, but in the "anthropological mechanisms of adoption"—who needs to be convinced, how to change standard treatment protocols, and how to help patients understand new metrics.

Theme 1: Cancer Diagnostics — AI Has Surpassed Clinical Standards, but Adoption Mechanisms Remain the Bottleneck

Barzilay points out that the current clinical standard for breast cancer risk is still based on the "breast density" classification method proposed in 1967 by a radiologist using the naked eye, whereas deep learning models can already "systematically identify patterns" to predict the risk of developing cancer within 1–5 years, with accuracy far exceeding existing standards.

Historical Context: In 1967, radiologist Wolf reviewed images of already diagnosed women, summarized several patterns by visual inspection, which were later encoded into four density categories. In 2019, this assessment became federal law — women are informed whether they fall into the "high density" (high risk) category. However, 40–50% of women are classified as high density. "If half the population is at high risk, this indicator is almost useless."
Data Chain: Barzilay emphasizes that algorithms are already "orders of magnitude better than current clinical practice," but "it's not an algorithm problem — the obstacle is that underexplored anthropological link."
Deduction and Falsification: She argues that transitioning from the density standard to an AI model requires answering: "Who needs to be convinced? How many hospitals need to run trials? How do we explain to women across the U.S. that this is a better indicator?" — Falsification condition: If, within the next 5–10 years, the FDA or major healthcare systems still fail to adopt AI risk models as the standard, it would indicate that institutional bottlenecks are harder to overcome than technical ones.

Barzilay cites the book American Sickness (Elizabeth Rosenthal), noting that "the incentive system in the U.S. healthcare system is extremely complex." Computer scientists entering this field must understand "how to navigate this system to drive adoption."

Theme 2: Data Acquisition – Hospitals Hold Data but Lack Incentive to Share; Patients Should Own Their Data

Barzilay revealed that after deciding to pursue cancer research, it took her "two years to obtain any meaningful dataset" – the U.S. still has no publicly available modern mammography dataset.

Mechanism Breakdown: Hospitals legally hold patient data, but "if it's given to the wrong person, the hospital loses a lot; if it's given to you, the hospital as a legal entity gains little." She believes the ideal solution is a "data donation" mechanism akin to organ donation – patients can choose whether to donate their data for research during medical visits.
Technical Approach: MIT is developing "learning on encoded data" – images are locally encoded before training the network, returning only results without exposing raw data. The natural language processing community is also improving de-identification techniques.
Personal Case: Barzilay disclosed that she receives mammograms from patients mailed to her MIT office almost weekly, asking her to run her model for a judgment – because patients "need information to make life-changing decisions, but the system doesn't provide it."
Inference: She argues that patient ownership of data is the right direction, but both Google Health and Microsoft Health Vault have shut down, indicating "either regulatory pressure, a lack of a business case, or resistance from hospitals."

Theme 3: Drug Design — ML Has Not Yet Been Involved in Any Approved Drug, But This Is "the Most Technically Exciting Area"

Barzilay notes that "no drug has been developed by an ML model, nor has ML played a significant role in any" — yet drug design is "a truly interesting and exciting open field from a technical standpoint."

Current process: Pharmaceutical companies identify candidate molecules through high-throughput screening (actually testing hundreds of thousands of compounds in the lab), and then chemists manually optimize them based on domain knowledge — adjusting solubility, reducing toxicity, and minimizing side effects. Barzilay comments: "Chemists have a very deep understanding of the domain; it's not observation-driven, there is a lot of science behind it."
ML entry points: ① Property prediction — inputting molecular structures (graphs) to predict properties such as biological activity; ② Molecular generation — similar to machine translation, encoding molecules into a latent space and then decoding improved versions. She emphasizes that "more powerful and interesting models indeed perform better, so there is room for ML innovation here."
Data chain: The cost of the traditional process is "prohibitively high," many diseases have no drugs at all, and failures of neurodegenerative disease drugs in late-stage trials are "financially catastrophic."
Inference: Barzilay argues that ML can learn from "millions of molecules and reactions," while "even the most experienced chemist cannot hold that much information in memory."

Theme 4: The Evolution of NLP — Statistical Methods Dominate, the Role of Linguistics "Greatly Diminished"

Barzilay recalled her entry into NLP in 1997: "Half the papers were rule-based, the other half were the first corpus-based papers — very simple, collect some statistics and make predictions."

Historical Context: Her first paper in 1997 "had no formulas, no evaluation, only output examples — that was the standard at the time." Since then, statistical methods have dominated the field, but "the role of linguistics has greatly diminished — today, flipping through conference proceedings, you can hardly find one or two papers that make interesting linguistic references."
Key Analogy: She compared the evolution of NLP to a problem in biology — linguistics attempted to "understand" language through structured representations like syntactic trees, but "this path didn't get very far"; statistical models, on the other hand, made significant progress by "using information from other sources."
Current Limitations: Machine translation performs excellently on news data, but "recipe translation is a disaster" — because recipes are not in the training set. She believes the core issue is "distribution shift" — when the training and test distributions differ, the system becomes extremely fragile.
Inference: She argues that "truly learning from a small number of samples" is the biggest unsolved problem in NLP today — "we write in papers about improving from 55% to 65%, but none of these methods are actually usable in practice."

Theme 5: The Turing Test and Human Cognition — Machines Do Not Need to "Understand" Like Humans

Barzilay takes a pragmatic stance on the Turing Test: machines do not need to understand language the way humans do, as long as they can reliably complete tasks.

Key analogy: "A calculator performs calculations in a way completely different from yours, yet it is highly effective — and that is fine. If we can accomplish certain tasks with high precision, we do not necessarily have to understand them in the same way we do."
Personal observation: She notes that Tencent's chatbot has "a large number of users chatting with it for hours," and recounts the MIT ELIZA story — secretaries would spend hours conversing with a program that "only did string matching and was very rudimentary." She concludes: "The issue is not how good the technology is, but how willing we are to believe it provides what we want."
Extrapolation: She argues that whether current approaches can pass a strict Turing Test (conversing with a neural network for an hour) is "a data problem, and also a problem of how it is trained — the system's true generalization and compositional abilities are limited under current capabilities."

Mentioned Positions

This section is not applicable — this issue is an academic/technical discussion and does not involve specific investable positions.

Judgments Worth Remembering

1. Barzilay believes AI cancer diagnosis algorithms are already "orders of magnitude" better than clinical standards, but the adoption bottleneck is not technical—the breast density standard (proposed in 1967) remains a federally mandated metric to inform women, yet 40-50% of women fall into the "high-risk" category, making the metric nearly useless. Falsification condition: If major healthcare systems have not adopted AI risk models within 5-10 years, institutional bottlenecks will prove harder to overcome than technical ones.

2. It took Barzilay two years to obtain any meaningful dataset—there is no publicly available modern mammography dataset in the U.S. She believes patients should own their data, but Google Health and Microsoft Health Vault have both shut down, suggesting "either regulatory pressure, no business case, or hospital resistance."

3. "No drug has been developed by an ML model"—drug design is "the most technically exciting open frontier." The current process relies on chemists manually optimizing, with costs "prohibitively high," while ML can learn "millions of molecules and reactions"—something even the most experienced chemist cannot do.

4. Barzilay proposes the "calculator analogy": Machines do not need to understand language the way humans do—"a calculator performs calculations in a way completely different from you, but it is extremely effective." She considers the demand for machines to "understand like humans" as "naive."

5. The biggest unsolved problem in NLP is "truly learning from small samples"—papers claiming few-shot methods that improve from 55% to 65% "none are actually usable in practice."

6. Barzilay cites the ELIZA story to illustrate the psychological dimension of the Turing test: MIT secretaries would spend hours chatting with a program that "only does string matching"—"the problem is not how good the technology is, but how willing we are to believe it provides what we want."

7. Barzilay believes personal mission should be independent of external recognition: "When I was young, I was mainly driven by external stimuli—to achieve this, to become that. Now much of my work is driven by 'what matters to me,' independent of external recognition." She thinks "vanity is everywhere, but at MIT there are also different forms of vanity."

8. She recommends two books: The Emperor of All Maladies (revealing "how imprecise and imperfect" the cancer drug development process is) and American Sickness (explaining the "complex incentive system" of the U.S. healthcare system), arguing that computer scientists entering healthcare must understand these.