I have a lot to say about this but I will keep it short. First, I think you’re underselling the insight that cancer isn’t a single disease (the Atlantic headline was shitty; of course cancer is a disease). This wasn’t obvious a priori. The fact that every case of cancer is a unique and horrible snowflake means that we can’t expect “a cure for cancer” any more than we can expect “a cure for car trouble”. You’re right, however, that some things are more likely to go wrong than others, and routine sequencing of tumors from each individual patient can help identify which treatments are most likely to help.
Second, I think there’s a link between the decrease in death rate from heart disease and the minimal death rate decline from cancer even with all the increased testing and new treatments. As they say, “something’s gonna kill ya,” and in my opinion dying from cancer at 65 instead of from a heart attack at 55 is still a win. As a comparator in a disease area where no treatments have really worked to date, see the death rates from Alzheimer’s disease.
Third, I’m as bullish on cancer immunotherapy as the next guy, but it turns out that many tumors produce an immunosuppressive environment, where T cells and NK cells just don’t do their thing very well. You can immunize against mutated protein fragments presented by the MHC all you want, but in an immunosuppressive microenvironment I still don’t think you’ll see those sweet sweet CRs.
Finally, even with all the regulatory barriers and misaligned incentives, pharma companies are still working on the best cancer therapy targets we know about. We (I work in pharma, so it’s “we”) certainly haven’t hung up the “Mission Accomplished” banner and moved on. While I expect continued insights about basic cancer biology to come from academic labs that receive public funding, future therapies will continue to arrive primarily from the private sector. The potential pecuniary reward for even incremental increases in cancer survival rate is high enough to keep key players interested.
You’re right, however, that some things are more likely to go wrong than others, and routine sequencing of tumors from each individual patient can help identify which treatments are most likely to help.
It can also help with amputating parts of people’s body for no useful medical purpose.
Second, I think there’s a link between the decrease in death rate from heart disease and the minimal death rate decline from cancer even with all the increased testing and new treatments. As they say, “something’s gonna kill ya,” and in my opinion dying from cancer at 65 instead of from a heart attack at 55 is still a win. As a comparator in a disease area where no treatments have really worked to date, see the death rates from Alzheimer’s disease.
This does suggest that cancer treatments aren’t completely worthless. The fact that the US manages compared to other OECD countries relatively low lifespan, high cancer survival rate and high cancer death rate however still points to goodharting to be responsible for a large part of the high cancer survival rate among OECD countries.
Third, I’m as bullish on cancer immunotherapy as the next guy, but it turns out that many tumors produce an immunosuppressive environment, where T cells and NK cells just don’t do their thing very well.
There’s a difference between natural NK cells and NK cells that gene manipulate to avoid features of the immunosuppressive enviroment. If a NK cells assumes that a cancer cell should have some tumor marker but the cancer cells of a particular patient don’t have that particular tumor marker you can alter the NK cells to not care about that tumor marker.
At current technology we can’t sequence a tumor and know what we have to change in the NK cells to make them work in the particular immunosuppressive environment of a particular patient for the NK cell to be operational in that enviroment, and then change them. It’s still technology that we can build.
Finally, even with all the regulatory barriers and misaligned incentives, pharma companies are still working on the best cancer therapy targets we know about.
One of my main points was that working on cancer therapy targets is not where most of the attention should go but on underlying technology platforms.
DeepMind beating Big Pharma at protein folding prediction suggests relatively little investment in the basic technology. It’s quite plausible that you actually need working protein folding prediction to do what I pointed towards above for modifying the NK cells to work in more immunosuppressive environments.
While I expect continued insights about basic cancer biology to come from academic labs that receive public funding, future therapies will continue to arrive primarily from the private sector.
I’m completely okay with the private sector developing therapies whenever they think that they can develop a working therapy and bring it profitably to market. My post was more about how the large chunk of government and non-profit money should be spent.
If would also be okay with simply redirected the cancer research budget elsewhere (like anti-aging) and leaving the problem completely to the private sector but there’s political desire to use public funds to do something about aging.
Thanks for engaging! I think there’s a real debate to be had about how public research money is spent. I put a higher expected value on continuing to fund basic cancer research than I think you do. I also am more bullish on doing working at the object level (going after specific targets) relative to the meta level (technology platforms). Maybe this is myopia on my part, working as I do in the pharmaceutical industry, but I have also spent a fair amount of time thinking about the problem.
DeepMind beating Big Pharma at protein folding prediction suggests relatively little investment in the basic technology.
I actually think DeepMind is plausibly the only entity in the world who could have made AlphaFold when they did. The sheer amount of compute involved puts it out of the reach of nearly everyone else, plus pharma companies would have found it hard to hire away the caliber of ML talent DeepMind attracts. There’s a case to be made that this is a nearly-ideal outcome for the pharma industry: the problem was cracked, publicly, by a company with little to no interest in making medicines. My prediction is that DeepMind either licenses the technology to pharma companies or contracts with them on specific targets (if the compute requirement is prohibitive for licensing). That seems to satisfy the incentives of DeepMind (this should be a significant money-maker, plus good publicity if and when their structures help lead to new drugs) and pharma (get structures for important targets that we can’t get other ways).
Solving protein folding doesn’t only give you the ability to know how existing proteins fold. It also gives you the ability to design new proteins.
If you take a problem like the one of NK cells and cancer immunity against them there will be cases where the machinery that NK cells have by human nature won’t be enough and you need to design new proteins for them to properly recognize the cancer cells.
You won’t get there by just licensing some protein folding technology from Google.
The sheer amount of compute involved puts it out of the reach of nearly everyone else, plus pharma companies would have found it hard to hire away the caliber of ML talent DeepMind attracts.
Compute and programmers are something you can hire. If the big pharma companies would be functional, it would be appropriate for each of them to spend a billion per year on AI.
A few years ago I spoke a few times with someone doing new business development at Pfizer. According to his perspective Pfizer is too bureaucratic to effective develop software.
It might be that new biotech companies that can actually manage to employ programmers in a productive way without stiffling them with too much bureaucracy will win out.
While not in the cancer area Proteon Pharmaceuticals would be a company where bioinformatics is at the core of their phage therapy product. As they grow and slowly push out most of the antibiotics from the market they have a need for more IT investment to better simulate how changes in phages and hosts interact and which mutations have which effects.
It might be that after they have a really good system to simulate phages, they go and reapply their knowledge to NK cells.
A company like Moderna or BioNTech that focus on cancer vaccines might also make major IT investments to optimize target selection and fully simulate the cancer cells to know which antigens they actually present.
Solving protein folding doesn’t only give you the ability to know how existing proteins fold. It also gives you the ability to design new proteins.
I don’t agree with this claim. AlphaFold gives you the ability to calculate how a given amino acid sequence is likely to fold. That is very different from being able to predict an amino acid sequence that performs a specific function or even has a given shape. Small modifications of known shapes or functionalities would be tractable using AlphaFold’s technology, but there are other ways to get that, for example directed evolution. Search in the space of amino acid sequences is possible in principle but even with several orders of magnitude increase in compute the size of the search space still seems intractable to me.
If the big pharma companies would be functional, it would be appropriate for each of them to spend a billion per year on AI.
Isn’t this significantly more than DeepMind spends? I realize increased competition for ML talent would drive up salaries but I just can’t see that kind of budget allocation happening for something that pharma companies don’t consider to be core to their business.
Thanks again for engaging. It’s been fun to see how someone in the Silicon Valley mindset looks at the biopharma landscape.
AlphaFold gives you the ability to calculate how a given amino acid sequence is likely to fold. That is very different from being able to predict an amino acid sequence that performs a specific function or even has a given shape. Small modifications of known shapes or functionalities would be tractable using AlphaFold’s technology, but there are other ways to get that, for example directed evolution.
This is basically why you need more then just licensing the technology from AlphaFold. You actually need to employ a bunch of programmers for a few years. It’s a hard problem but it’s solveable.
Isn’t this significantly more than DeepMind spends? I realize increased competition for ML talent would drive up salaries but I just can’t see that kind of budget allocation happening for something that pharma companies don’t consider to be core to their business.
Tesla is today worth more then the other car makers combined. The legacy car makers largely understood to late that driverless cars are going to be a core part of their business.
It might be that currently the idea of some big pharma companies is to wait for the equivalent of Cruise and then buy it.
Besides understanding what goes on in a cell with computer models and protein design, predicting phase 3 trial outcomes better based on phase 1 data would be another task where a lot of data can be gathered and analysed with machine learning. There’s an incredible amount of profit in prediction phase 3 trial outcomes before doing the trial.
I have a lot to say about this but I will keep it short. First, I think you’re underselling the insight that cancer isn’t a single disease (the Atlantic headline was shitty; of course cancer is a disease). This wasn’t obvious a priori. The fact that every case of cancer is a unique and horrible snowflake means that we can’t expect “a cure for cancer” any more than we can expect “a cure for car trouble”. You’re right, however, that some things are more likely to go wrong than others, and routine sequencing of tumors from each individual patient can help identify which treatments are most likely to help.
Second, I think there’s a link between the decrease in death rate from heart disease and the minimal death rate decline from cancer even with all the increased testing and new treatments. As they say, “something’s gonna kill ya,” and in my opinion dying from cancer at 65 instead of from a heart attack at 55 is still a win. As a comparator in a disease area where no treatments have really worked to date, see the death rates from Alzheimer’s disease.
Third, I’m as bullish on cancer immunotherapy as the next guy, but it turns out that many tumors produce an immunosuppressive environment, where T cells and NK cells just don’t do their thing very well. You can immunize against mutated protein fragments presented by the MHC all you want, but in an immunosuppressive microenvironment I still don’t think you’ll see those sweet sweet CRs.
Finally, even with all the regulatory barriers and misaligned incentives, pharma companies are still working on the best cancer therapy targets we know about. We (I work in pharma, so it’s “we”) certainly haven’t hung up the “Mission Accomplished” banner and moved on. While I expect continued insights about basic cancer biology to come from academic labs that receive public funding, future therapies will continue to arrive primarily from the private sector. The potential pecuniary reward for even incremental increases in cancer survival rate is high enough to keep key players interested.
It can also help with amputating parts of people’s body for no useful medical purpose.
This does suggest that cancer treatments aren’t completely worthless. The fact that the US manages compared to other OECD countries relatively low lifespan, high cancer survival rate and high cancer death rate however still points to goodharting to be responsible for a large part of the high cancer survival rate among OECD countries.
There’s a difference between natural NK cells and NK cells that gene manipulate to avoid features of the immunosuppressive enviroment. If a NK cells assumes that a cancer cell should have some tumor marker but the cancer cells of a particular patient don’t have that particular tumor marker you can alter the NK cells to not care about that tumor marker.
At current technology we can’t sequence a tumor and know what we have to change in the NK cells to make them work in the particular immunosuppressive environment of a particular patient for the NK cell to be operational in that enviroment, and then change them. It’s still technology that we can build.
One of my main points was that working on cancer therapy targets is not where most of the attention should go but on underlying technology platforms.
DeepMind beating Big Pharma at protein folding prediction suggests relatively little investment in the basic technology. It’s quite plausible that you actually need working protein folding prediction to do what I pointed towards above for modifying the NK cells to work in more immunosuppressive environments.
I’m completely okay with the private sector developing therapies whenever they think that they can develop a working therapy and bring it profitably to market. My post was more about how the large chunk of government and non-profit money should be spent.
If would also be okay with simply redirected the cancer research budget elsewhere (like anti-aging) and leaving the problem completely to the private sector but there’s political desire to use public funds to do something about aging.
Thanks for engaging! I think there’s a real debate to be had about how public research money is spent. I put a higher expected value on continuing to fund basic cancer research than I think you do. I also am more bullish on doing working at the object level (going after specific targets) relative to the meta level (technology platforms). Maybe this is myopia on my part, working as I do in the pharmaceutical industry, but I have also spent a fair amount of time thinking about the problem.
I actually think DeepMind is plausibly the only entity in the world who could have made AlphaFold when they did. The sheer amount of compute involved puts it out of the reach of nearly everyone else, plus pharma companies would have found it hard to hire away the caliber of ML talent DeepMind attracts. There’s a case to be made that this is a nearly-ideal outcome for the pharma industry: the problem was cracked, publicly, by a company with little to no interest in making medicines. My prediction is that DeepMind either licenses the technology to pharma companies or contracts with them on specific targets (if the compute requirement is prohibitive for licensing). That seems to satisfy the incentives of DeepMind (this should be a significant money-maker, plus good publicity if and when their structures help lead to new drugs) and pharma (get structures for important targets that we can’t get other ways).
Solving protein folding doesn’t only give you the ability to know how existing proteins fold. It also gives you the ability to design new proteins.
If you take a problem like the one of NK cells and cancer immunity against them there will be cases where the machinery that NK cells have by human nature won’t be enough and you need to design new proteins for them to properly recognize the cancer cells.
You won’t get there by just licensing some protein folding technology from Google.
Compute and programmers are something you can hire. If the big pharma companies would be functional, it would be appropriate for each of them to spend a billion per year on AI.
A few years ago I spoke a few times with someone doing new business development at Pfizer. According to his perspective Pfizer is too bureaucratic to effective develop software.
It might be that new biotech companies that can actually manage to employ programmers in a productive way without stiffling them with too much bureaucracy will win out.
While not in the cancer area Proteon Pharmaceuticals would be a company where bioinformatics is at the core of their phage therapy product. As they grow and slowly push out most of the antibiotics from the market they have a need for more IT investment to better simulate how changes in phages and hosts interact and which mutations have which effects.
It might be that after they have a really good system to simulate phages, they go and reapply their knowledge to NK cells.
A company like Moderna or BioNTech that focus on cancer vaccines might also make major IT investments to optimize target selection and fully simulate the cancer cells to know which antigens they actually present.
I don’t agree with this claim. AlphaFold gives you the ability to calculate how a given amino acid sequence is likely to fold. That is very different from being able to predict an amino acid sequence that performs a specific function or even has a given shape. Small modifications of known shapes or functionalities would be tractable using AlphaFold’s technology, but there are other ways to get that, for example directed evolution. Search in the space of amino acid sequences is possible in principle but even with several orders of magnitude increase in compute the size of the search space still seems intractable to me.
Isn’t this significantly more than DeepMind spends? I realize increased competition for ML talent would drive up salaries but I just can’t see that kind of budget allocation happening for something that pharma companies don’t consider to be core to their business.
Thanks again for engaging. It’s been fun to see how someone in the Silicon Valley mindset looks at the biopharma landscape.
This is basically why you need more then just licensing the technology from AlphaFold. You actually need to employ a bunch of programmers for a few years. It’s a hard problem but it’s solveable.
Tesla is today worth more then the other car makers combined. The legacy car makers largely understood to late that driverless cars are going to be a core part of their business.
It might be that currently the idea of some big pharma companies is to wait for the equivalent of Cruise and then buy it.
Besides understanding what goes on in a cell with computer models and protein design, predicting phase 3 trial outcomes better based on phase 1 data would be another task where a lot of data can be gathered and analysed with machine learning. There’s an incredible amount of profit in prediction phase 3 trial outcomes before doing the trial.