Over the past few weeks, I’ve been transfixed, and saddened, by Eric Gilliam’s three-part series about the history of MIT (my alma mater). I’ll post a few quotations and responses below, but if you’re interested you should just go read the original essays
(1,
2,
3).
Why MIT Was Created
Professors who are not steeped in hands-on industrial practice could not produce the kinds of workers that were immediately useful to industry. These schools were outputting the kind of men that [Thomas] Edison, and many others mentioned above, did not believe were meeting the needs of industry. And the technical know-how taught in trade schools was great, but an ideal institute of technology should also impart some higher engineering and scientific knowledge to students to enable them to be more innovative, intelligent problem-solvers.
So, MIT was founded to solve this problem. This school was not designed to be a place for purely lecturing and rote learning. A smattering of intelligent men from industry and university men with an applied bent to them made up the original faculty. Content was lectured as needed, but what differentiated MIT was its innovative use of the laboratory method. Instructors taught “through actual handling of the apparatus and by working on problems, shoulder to shoulder with the boys.” And the schedule, from 9-5 (with a lunch break) 5 days a week and additional class on Saturday was meant to simulate a normal work schedule and, thus, ease the eventual transition to life in the working world.
(part 1)
This quote highlights how MIT was intended to be a counter-cultural university, founded on a distinctly different model than other institutions (like Harvard).
MIT was not meant to be a center of learning and theoretical research, but a school focusing on training the next generation of industrial leaders.
How MIT Supported Itself
But [MIT President] Maclaurin had an idea: self-support. MIT would capitalize on its own assets and earn money by formally offering its services to industry on a larger scale. High numbers of industrial partners had been eager to engage in ad-hoc courses of research with MIT’s applied professors, often paid for by the company, anyway. Why not turn this into a much larger, more formal program that was facilitated by the Institute? The idea would grow into what was known as the Technology Plan.
(part 2)
MIT operated on a different funding model than other universities, relying on support from industry.
This is, in essence, what I proposed several weeks ago in my reflection on the similarities between graduate school and consulting.
This was seen as important and honorable by its leaders at the time:
“There could be no more legitimate way for a great scientific school to seek support than by being paid for the service it can render in supplying special knowledge where it is needed... Manufacturers may come to us with problems of every kind, be they scientific, simple, technical or foolish. We shall handle each seriously, giving the best the institute has at its disposal” - William Walker, head of the Division for Industrial Cooperation and Research
(part 2)
Why MIT Changed Paths
The answer to this question is the subject of Gilliam's
third post.
It's a bit too complex to fully summarize here, but there were a few key factors:
The massive increase in post-WWII funding for science allowed MIT to shift away from applied research and towards basic research.
The growing importance of industrial research labs at inter alia Bell Labs, GE, and DuPont meant that the need for academic applied research was decreasing, and also that academically trained researchers were valuable assets to industry. (This was the case only in the post-war period, however, and by 1980 the era of large industrial research labs had mostly concluded.)
A growing distaste for pursing "boundary work" that could also be done in industry, as opposed to pursuing a uniquely academic niche.
Crucially, the first two factors are less true today than they were when MIT made this decision, implying that the niche filled by "Old MIT" could be occupied again today.
Why A School Like Old MIT Should Still Exist
It seems clear, given MIT’s transition to a more university style of education, that we are left with a hole. We do not have an elite hybrid technical school/applied research institute like this that can draw top talent away from places like Harvard and Stanford to its more hands-on style of education. But, as a country where the manufacturing sector is shrinking (and median wages aren’t doing so well either), we may need a new MIT now more than ever.
There are plenty of individuals at top schools who COULD be swayed to attend a place like this. Speaking for Stanford, where I went to undergrad, there was a large population of people who majored in mechanical engineering and were disenchanted because they did almost exclusively problem set work and very little building of anything real. And I knew even more people majoring in other subjects who abandoned mechanical engineering and majors like it for this reason! “We’re training you to be mechanical engineering managers, not traditional mechanical engineers,” was a common line used in the department. And, while that is a fine goal for a program, it is not what many of the students seem to want. What if I just want to be a top-flight regular engineer who can build awesome stuff?
(part 1)
There can and should be individuals who spend almost all of their time on pure research. But it is probably bad for future progress to allow too many of these individuals to work in an environment in which few of their peers are spending a substantial amount of time working on industrial applications and problems. No matter what, some basic research will always find a way of trickling its way down into practical industrial importance. But allowing pure researchers to be siloed from the acquaintance of those who work on industrial applications — and not just the need to work on those problems themselves — feels like it is setting the system up for inefficiency. When we look back on the era of explosive productivity in areas of basic research like physics and math in the early 1900s, even the purest of pure researchers at the time tended to have regular interactions either with industry or with researchers who did industry-related research — due to industry contracts themselves, close friends who did industry work regularly, or conscription to work on military.
(part 2)
Gilliam's conclusions seem broadly correct to me. While MIT is still a great school, it's no longer pursuing a distinct model for education. The main factors distinguishing MIT from peer institutions are cultural, and even those are being activelysuppressed by the current administration.
In total it took less than a century for the entrepreneurial mindset of MIT, a "startup university", to be replaced by the exact institutional conservatism it was founded to oppose. "You either die a hero or live long enough to see yourself become the villain."
More broadly, there's a broad sense today that innovation, especially in the physical world, is slowing (Peter Thiel may be the most notable proponent of this claim).
A century ago, Americans could build whole networks of subways with comparatively primitive technology; now, something as simple as building a single subway station has become a Herculean task.
I don't mean to draw too direct of a causal connection between the end of Old MIT and the decline in real-world innovation, but perhaps a new school focused on unglamorous, nuts-and-bolts innovation rather than holistic education is exactly what the US needs now.
Organic chemists often think in terms of potential energy surfaces, especially when plotting the results of a computational study.
Unfortunately it is non-trivial to generate high-quality potential energy surfaces. It's not too difficult to sketch something crude in ChemDraw or Powerpoint, but getting the actual
barrier heights correct and proportional has always seemed rather tedious to me.
I've admired the smooth potential energy surfaces from the Baik group for years, and so several months ago I decided to try and write my own
program to generate these diagrams. I initially envisioned this as a python package (with the dubiously clever name of pypes), but it turned out to be simpler than expected, such that I
haven't actually ever turned it into a library. It's easier to just copy and paste the code into various Jupyter notebooks as needed.
Here's the code:
# get packages
import numpy as np
import scipy.interpolate as interp
import matplotlib.pyplot as plt
# make matplotlib look good
plt.rc('font', size=11, family="serif")
plt.rc('axes', titlesize=12, labelsize=12)
plt.rc(['xtick', 'ytick'], labelsize=11)
plt.rc('legend', fontsize=12)
plt.rc('figure', titlesize=14)
%matplotlib inline
%config InlineBackend.figure_format='retina'
# x and y positions. y in kcal/mol, if you want, and x in the range [0,1].
Y = [2.49, 3.5, 0, 20.2, 19, 21.5, 20, 20.3, -5]
X = [0, 0.15, 0.3, 0.48, 0.55, 0.63, 0.70, 0.78, 1]
# labels for points. False if you don't want a label
label = ["label1", False, "label2", "label3", "label4", "label5", "label6", "label7", "label8"]
#### shouldn't need to modify code below this point too much...
# autodetect which labels correspond to transition states
TS = []
for idx in range(len(Y)):
if idx == 0 or idx == len(Y)-1:
TS.append(False)
else:
TS.append((Y[idx] > Y[idx+1]) and (Y[idx] > Y[idx-1]))
# sanity checks
assert len(X) == len(Y), "need X and Y to match length"
assert len(X) == len(label), "need right number of labels"
# now we start building the figure, axes first
f = plt.figure(figsize=(8,8))
ax = f.gca()
xgrid = np.linspace(0, 1, 1000)
ax.spines[['right', 'bottom', 'top']].set_visible(False)
YMAX = 1.1*max(Y)-0.1*min(Y)
YMIN = 1.1*min(Y)-0.1*max(Y)
plt.xlim(-0.1, 1.1)
plt.tick_params(axis='x', which='both', bottom=False, top=False, labelbottom=False)
plt.ylim(bottom=YMIN, top=YMAX)
ax.plot(-0.1, YMAX,"^k", clip_on=False)
# label axes
plt.ylabel("Gibbs Free Energy (kcal/mol)")
plt.xlabel("Reaction Coordinate")
# plot the points
plt.plot(X, Y, "o", markersize=7, c="black")
# add labels
for i in range(len(X)):
if label[i]:
delta_y = 0.6 if TS[i] else -1.2
plt.annotate(
label[i],
(X[i], Y[i]+delta_y),
fontsize=12,
fontweight="normal",
ha="center",
)
# add connecting lines
for i in range(len(X)-1):
idxs = np.where(np.logical_and(xgrid>=X[i], xgrid<=X[i+1]))
smoother = interp.BPoly.from_derivatives([X[i], X[i+1]], [[y, 0] for y in [Y[i], Y[i+1]]])
plt.plot(xgrid[idxs], smoother(xgrid[idxs]), ls="-", c="black", lw=2)
# finish up!
plt.tight_layout()
plt.show()
The output looks like this:
The potential energy surface generated by the above code.
If you like how this looks, feel free to use this code; if not, modify it and make it better! I'm sure this isn't the last word in potential-energy-surface creation, but it's good enough for me.
Now that our work on screening for generality has finally been published in Nature, I wanted to first share a few personal reflections and then highlight the big conclusions that I gleaned from this project.
This project originated from conversations I had with Eugene Kwan back in February 2019, when I was still an undergraduate at MIT. Although at the time our skills were almost completely non-overlapping, we shared both an interest in “big data” and high-throughput experimentation and a conviction that organic chemistry could benefit from more careful thinking about optimization methods.
After a few months of work, Eugene and I had settled on the idea of a “catalytic reaction atlas” (in analogy to the cancer genome atlas) where we would exhaustively investigate catalysts, conditions, substrates, etc. for a single asymmetric reaction and then (virtually) compare different optimization methods to see which algorithms led to the best hits. Even with fairly conservative assumptions, we estimated that this would take on the order of 105 reactions, or about a year of continuous HPLC time, meaning that some sort of analytical advance was needed.
A slide comparing different optimization strategies, from April 2019. Multi-substrate screening is proposed as one of many different algorithms.
When I proposed this project to Eric, he was interested but suggested we focus more narrowly on the question of generality, or how to discover reactions with broad substrate scope. In an excited phone call, Eugene and I had the insight that we could screen lots of substrates at once by using mass spectrometry, thus bypassing our analytical bottleneck and enabling us to access the “big data” regime without needing vast resources to do so.1
Getting the analytical technology to work took about two years of troubleshooting. We were lucky to be joined by Spencer, an incredible analytical chemist and SFC guru, and eventually were able to get reproducible and accurate data by a combination of experimental insights (running samples at high dilution) and computational tweaks (better peak models and fitting algorithms). To make sure that the method was working properly, we ran validation experiments both on a bunch of scalemic samples and on a varied set of complex pharmaceutical racemates.
Choosing the proper reaction took a bit of thought, but once we settled on a set of substrates and catalysts the actual experiments were a breeze. Almost all the screening for this project was done in November–December 2021: in only a few hours, I could easily run and analyze hundreds of reactions per week.
I want to conclude by sharing three high-level conclusions that I’ve taken away from working on this project; for the precise scientific conclusions of this study, you can read the paper itself.
1. Chemical space is big, so how you search matters
There are a ton of potential catalysts waiting to be discovered, and it seems likely that almost any hit can be optimized to 90% ee by sufficient graduate-student hours. Indeed, one of the reasons we selected the Pictet–Spengler reaction was the diversity of different catalyst structures capable of giving high enantioselectivity. But just because you can get 90% ee from a given catalyst family doesn’t mean you should: it might be terrible for other substrates, or a different class of catalysts might be much easier to optimize or much more reactive.
Understanding how many catalysts are out there to be discovered should make us think more carefully about which hits we pursue, since our time is too valuable to waste performing needless catalyst optimizations. In this study, we showed that screening only one substrate can be misleading when the goal is substrate generality, but one might prefer to screen for other factors: low catalyst loading, tolerance of air or water, or recyclability all come to mind. In all cases, including these considerations in initial screens means that the hits generated are more likely to be relevant to the final goal. Just looking for 90% ee is almost certainly not the best way to find a good reaction.
2. Don’t ignore analytical chemistry
Although assay development is a normal part of many scientific fields, many organic chemists seem to barely consider analytical chemistry in their research. Any ingenuity is applied to developing new catalysts, while the analytical method remains essentially a constant factor in the background. This is true even in cases where the analytical workflow represents a large fraction of the project (e.g. having to remove toluene before NMR for every screen).
This shouldn’t be the case! Spending time towards the beginning of a project to develop a nice assay is an investment that can yield big returns: this can be as simple as making a GC calibration curve to determine yield from crude reaction mixtures, or as complex as what we undertook here. Time is too valuable to waste running endless columns.
More broadly, it seems like analytical advances (e.g. NMR and HPLC) have had a much bigger impact on the field than any individual chemical discoveries. Following this trend forward in time would imply that we should be making bigger investments in new analytical technologies now, to increase scientist productivity in the future.
3. A little computer science can go a long way
A key part of this project (mentioned only briefly in the paper) was developing our own peak-fitting software that allowed us to reliably fit overlapped peaks. This was computationally quite simple and relied almost entirely on existing libraries (e.g. scipy and lmfit), but took a certain amount of comfort with signal processing / data science.2 We later ended up moving our software pipeline out of unwieldy Jupyter notebooks and into a little Streamlit web app that Eugene wrote, which allowed us to quickly and easily get ee values from larger screens.
Neither of these two advances required significant coding skill; rather, just being able to apply some computer science techniques to our chemistry problem unlocked new scientific opportunities and massive time savings (a la Pareto principle). Moving forward, I expect that programming will become a more and more central tool in scientific research, much like Excel is today. Being fluent in both chemistry and CS is currently a rare and valuable combination, and will only grow in importance in the coming decades.
Thanks to Eugene Kwan for reading a draft of this post.
Footnotes
I'd like to propose the following principle: any sufficiently clever analytical technique inevitably depends on mass spectrometry. If you don't believe me, just look at the field of proteomics...
One common misconception in mechanistic organic chemistry is that reactions are accelerated by speeding up the rate-determining step.
This mistaken belief can lead to an almost monomaniacal focus on determining the nature of the rate-determining step.
In fact, it's more correct to think of reactions in terms of the rate-determining span: the difference between the resting state and the highest-energy transition state.
(I thank Eugene Kwan's notes for introducing me to this idea.)
In this post, I hope to demonstrate the veracity of this concept by showing that, under certain idealized assumptions, the existence of a low-energy intermediate has no effect on rate. Consider the following system:
Concerted and stepwise mechanisms.
We can imagine plotting these two mechanisms on a potential energy surface:
Concerted and stepwise mechanisms, on a PES.
In this example, X = Y + Z; the energy of the transition state and ground state are the same in both cases, and only the presence (or absence) of an intermediate differentiates the two potential energy surfaces. We will now compute the rate of product formation in both cases.
Using the Eyring–Polyani equation, it's straightforward to arrive at an overall rate for the concerted reaction as a function of the barrier:
k = kBT/h * exp(-X/RT)
rateconcerted = k * [SM]
rateconcerted = kBT/h * exp(-X/RT) * [SM]
The stepwise case is only slightly more complicated. Assuming that the barrier to formation of the intermediate is much lower than the barrier to formation of the product, and that the intermediate is substantially lower in energy than the rate-limiting transition state, we can apply the pre-equilibrium approximation:
ratestepwise = k2 * [INT]
k2 = kBT/h * exp(-Z/RT)
ratestepwise = kBT/h * exp(-Z/RT) * [INT]
Solving for [INT] is straightforward, and we can plug the result in to get our final answer:
As promised, the rates are the same—where the preequilibrium approximation holds, the existence of an intermediate has no impact on rate.
All that matters is the relative energy of the transition state and the ground state.
This method of thinking is particularly useful for rationalizing tricky Hammett trends. For instance, it's known that electron-rich indoles react much faster in Brønsted-acid-catalyzed Pictet–Spengler reactions, even though these reactions proceed through rate-determining elimination from a carbocation. Since electron-poor carbocations are more acidic, simple analysis of the rate-determining step predicts the opposite trend.
However, if we ignore the intermediate, it's clear that the transition state contains much more carbocationic character than the ground state, and so electron-donating groups will stabilize the transition state relative to the ground state and thereby accelerate the reaction. Thinking about intermediates is a great way to get confused; to understand trends in reactivity, all you need to consider is the transition state and the ground state.
The growing accessibility of computational chemistry has, unfortunately, led to a preponderance of papers with bad computations. Organic chemists are all too familiar with the “DFT section” of an otherwise high-quality publication which typically contains a transition-state structure or two, some sort of enigmatic cartoon purporting to explain the observed selectivity, and perhaps an uninterpretable NCIPLOT cited as evidence for the preceding claims.1
Faced with this sort of landscape, experimental chemists typically adopt one of two faulty heuristics: excessive credulity or universal skepticism. Being too trusting is dangerous, as evidenced by work showcasing the manifold ways that simulations can deceive the unwary scientist. Almost anyone who’s made a catalyst predicted to be better by computations knows this well (even when the computations are your own).
However, equally dangerous—and, in my view, less appreciated—is the creeping ennui that diminishes the entire field. This is exemplified by statements like “I don’t believe computations can ever be predictive,” “You can make DFT say anything you want to,” or, more delicately, “Computations are more for generating hypotheses, not being physically correct.” Although most people may be too polite to admit this to their computational collaborators, this nihilism is pervasive—just listen to the conversations as students walk back from a departmental seminar.
This viewpoint is wrong. The existence of bad computational models does not mean that all models are bad, nor does it imply that the task of creating models is inherently futile. Examples from other scientific fields, like orbital mechanics and fluid dynamics, indicate that computations can achieve impressive degrees of accuracy and become pivotal and trustworthy components of the scientific process. Closer to home, even the most skeptical chemists would admit that for e.g. calculating IR frequencies in the ground state, DFT shows impressive predictive accuracy (modulo the usual systematic error). There’s no intrinsic reason why accurately modeling chemical systems, even prospectively, ought to be impossible; chemistry is not a social science.
Why, then, is this variety of skepticism so common? Part of the problem comes from the bewildering milieu of options available to practitioners in the field. While a seasoned expert can quickly assess the relative merits of BYLP/MIDI! and DSD-PBEP86/def2-TZVP, to the layperson it’s tough to guess which might be superior. Without transparent heuristics by which to judge the quality of computational results, it’s no surprise that zeroth-order approximations (“DFT good” or “DFT bad”) have become so common among non-experts.2
Another issue is the generally optimistic demeanor of computational chemists towards their field. While the temptation to emphasize the potential upside of one’s research area is understandable, overestimating the capabilities of state-of-the-art technology inevitably leads to a reckoning when the truth becomes obvious. Except in certain circumscribed cases, we are still far from any predictive models of reactivity or selectivity for typical solution-phase reactions, various purported “breakthroughs” notwithstanding. Based on questions I’ve heard in talks, this uncomfortable truth is not universally understood by experimental audiences.
What, then, are the practical conclusions for computational chemists? Firstly, we should not be afraid to be our field’s own harshest critics. Allowing low-quality work into the literature erodes trust in our field; although raising our standards may be difficult and unpopular in the short term (kinetics), in the long run it will benefit the field (thermodynamics). You never get a second chance at a first impression; every bad paper published causes good papers to get that much less attention.
Secondly, we should work to develop consistent standards and workflows by which one can obtain reliable computational results. Just like there are accepted means by which new compounds are characterized (1H NMR, 13C NMR, HRMS, IR), there ought to be transparent methods by which transition states can reliably be found and studied. The manifold diversity of parameters employed today is a sign of the field’s immaturity—in truly mature fields, there’s an accepted right way to do things.3 The growing popularity of tools like crest is an important step in this direction, as is the ability to to use high-level post-Hartree–Fock wavefunction methods like DLPNO-CCSD(T) to refine single-point energies.
Finally, we must be honest about the limitations of our techniques and our results. So much about the chemical world remains mysterious and far beyond our understanding, let alone our ability to reproduce in silico. Far from being a failure for the field, however, this is something to be acknowledged and celebrated; science is only possible when there remain secrets to be found.
Between the Scylla of gullible credulity and the Charybdis of defensive nihilism, we must chart a middle way.
Thanks to Hayden Sharma for reading a draft of this post.
Footnotes
NCIPLOT is a program which allows one to visualize non-covalent interactions; although the output can be useful, for large molecules it's also very overwhelming.
A related idea is "epistemic learned helplessness", where people unable to evaluate the quality of a certain kind of argument resolve not to be persuaded by this argument one way or the other.
I want to thank Frank Neese and coworkers for publishing their entire transition-state-finding workflow in detail, which I've found very useful and is certainly a step in the right direction.
