Ron Boger

Functional luciferase design with deep learning

2023-08-16T00:00:00+00:00

Written in late Mar 2023, a quick review of de-novo enzyme design with deep learning from David Baker’s group. Christian Dallago and Kevin Yang wrote a fantastic Nature News and Views piece a couple months later, describing the technique with some more depth.

A deep-learning-based “family-wide hallucination” approach generates idealized protein structures for designing artificial luciferases with high selectivity and thermostability, enabling the creation of highly active and specific biocatalysts with broad biomedical applications.

De novo enzyme design has been a grand challenge in biology, as it seeks to create tailor-made enzymes that overcome the limitations of naturally occurring ones. The complexity of protein sequence-structure relationships and the scarcity of suitable protein structures have made it a difficult task. However, breakthroughs in computational methods and deep learning approaches are starting to unlock the potential of de novo enzyme design, paving the way for the development of highly specific, efficient, and stable biocatalysts that can revolutionize various fields, including biomedicine, environmental remediation, and industrial biotechnology.

Luciferases are a particular enzyme of interest, as they catalyze the oxidation of luciferin substrates, producing bioluminescent light that is widely used in bioassays and biomedical imaging. Their ability to generate luminescent photons without the need for an external excitation light source results in higher sensitivity and reduced autofluorescence or phototoxicity in live animal models and biological samples. However, the development of luciferases as molecular probes faces several challenges, including the limited number of native luciferases identified, structural instability due to multiple disulfide bonds, poor recognition of synthetic luciferins with more desirable photophysical properties, and low substrate specificity, which hinders multiplexed imaging for tracking multiple processes in parallel. Building on a growing body of knowledge regarding protein structure prediction ¹, ² with deep learning and small molecule-protein docking ³, Yeh et al ⁴ chose to design luciferases. The authors show that they can leverage deep learning to design luciferases that are small, highly stable, well-expressed in cells, substrate-specific, and need no cofactors to function.

Yeh and colleagues chose a synthetic luciferin, diphenylterazine (DTZ), as the target substrate because of its high quantum yield, red-shifted emission, favorable in vivo pharmacokinetics, and lack of required cofactors for light emission. Few native structures with favorable binding pockets to DTZ exist, so a computational screen was performed to identify a greater number of protein scaffolds. They computationally docked DTZ into 4,000 native small-molecule-binding proteins, and found that nuclear transport factor 2 (NTF2)-like folds have binding pockets with appropriate shape complementarity and size for DTZ placement.

Using the NTF2-like superfamily as the target topology, the authors selected 2,000 naturally occurring proteins within the superfamily. They sought to generate more proteins with the desired fold using “family wide hallucination”, since native NTF2 structures have features that are not ideal such as long loops that compromise stability. They conducted a Monte Carlo search in sequence space, guided by the predicted structure confidence and a topology-specific loss function over core residue pair geometries. This resulted in 1,615 hallucinated NTF2 scaffolds with greater shape-complementary binding pockets for DTZ than the native small-molecule-binding proteins, as measured by RIF binding energies.

A schematic representation of the family-wide hallucination approach to design a novel luciferase taken by Yeh et al.

Yeh and colleagues sought to design an effective active-site into their generated protein scaffolds. Using insights from quantum chemistry calculations ⁵ and experimental data ⁶, they designed a shape catalytic site that stabilizes the anionic state of DTZ and lowers the single-electron transfer (SET) energy barrier. To this end, they generated anionic DTZ conformers and used the RifGen ⁷, ⁸ method to generate rotamer interaction fields (RIFs) of preferred configurations of the scaffold amino acid residues based on their hydrogen-bonding and nonpolar interactions with DTZ. 50,000 docks with the most favorable interactions with DTZ were chosen for further sequence optimization using RosettaDesign ⁹. Interestingly, Yeh and colleagues used RosettaDesign to optimize the remainder of the sequence instead of ProteinMPNN ¹⁰, which was used in their work to design (h-CTZ)-specific luciferases. These designs were then filtered down to 7,648 proteins on the basis of ligand-binding energy, protein-ligand hydrogen bonds, shape complementarity, and contact molecular surface.

This protein library was cloned into E.Coli for expression and screened for bioluminescent activity. The activity of selected colony clones were confirmed by a 96-well plate expression. 3 active designs were identified; the top named LuxSit. Using circular dichroism (CD), the authors found an organized α-β structure for LuxSit. CD melting experiments further showed the protein was not fully unfolded at 95°C, indicating thermal stability.

LuxSit was further optimized using site saturation mutagenesis, in which each residue in the substrate-binding pocket was mutated to every other amino acid. This analysis led to further optimized LuxSit variants, dubbed LuxSit-i and LuxSit-f, with catalytic efficiencies comparable to native luciferases. The authors also purified LuxSit for size-exclusion chromatography which indicated that it is a single, uniform entity without forming any aggregates or multimeric complex. Yeh and colleagues then tested expression of LuxSit-i in live mammalian cells. They found that their engineered protein has high substrate specificity, an impressive feat compared to promiscuous native and previously engineered luciferases which act on many luciferin substrates.

Computational enzyme design has been constrained by the number of available scaffolds. Using deep learning to produce a large number of de-novo designed scaffolds, Yeh et al were able to design a novel luciferase with desirable biophysical characteristics for use as a biological probe. In conclusion, the development of the family-wide hallucination method has significant implications for protein engineering and catalysis research. By generating a vast number of de-novo-designed scaffolds, this approach opens up new possibilities for identifying potential substrate-binding and catalytic residues, even in cases where the reaction mechanism is not fully understood. The ability to explore a wide range of structural and catalytic hypotheses with shape and chemically complementary binding pockets, but different catalytic residue placements, could greatly enhance our understanding of enzyme function and lead to the design of novel and more efficient catalysts.

References

Anti-TechBio

2022-06-29T00:00:00+00:00

Cowritten with my friend Dennis Gong, who’s currently a Harvard-MIT PhD Student. Cross posted on Dennis’s site here.

The struggles of public biotech markets have led to sector wide reflection. Specifically, there has been a reckoning of TechBio businesses, such as Recursion, Ginkgo, Zymergen, and others. The first TechBio investments made in the mid 2010s are amongst the biggest losers in today’s public markets. Since 2010, TechBio businesses represent 5 of the 10 worst performing IPOs while accounting for just 5% of market entrants. Our view is that this collapse of vision has roots in naive forecasting and fundamental misunderstandings of the purpose and nature of utilizing biotechnologies for social benefit.

The thesis then as it is now, is familiar to most. At its core, it is the view that taking an engineering approach to problem solving is orders of magnitude more effective than whatever existing systems are being used. By applying new technologies, drug development can be faster, less expensive, and scalable. TechBio is engineering biology with a tech focused approach. These companies are typically platforms, aiming to have multiple shots on goal, instead of focusing on singular indications. It’s about controlling biological machinery, turning discovery into search, creating engines that continuously optimize until a drug pops out.

Several groups have written interesting articles about the upside and promise of TechBio. Those in venture are quite good at telling the story, building followings, and seeding tons and tons of companies. It’s an exciting time and everyone needs to know about it. How can you not be excited when…

“Bio is being digitized. It has become programmable.”

“The 21st century is for the biorevolution”

“Biology is eating the world”

The costs of drug development continue to climb, reaching up to $2 billion by some estimates. Meanwhile, increasing the size of drug R&D efforts has led to declining efficiencies, rather than more discovery, and new drug approvals have seemingly gotten less exciting – more me-too approvals with lower bars for efficacy and greater leniency for toxicity. It seems that amazing news happens every day in the biotech industry, but we can’t always be at the cusp. When will TechBio and all its promises arrive and deliver on new medicines?

The promises arise partly from silicon valley hubris, the idea that any industry should and will be disrupted by software and machine learning. They arise from real frustrations academics and industry scientists have with existing workflows of doing science. And they arise from a genuine optimism that everyone in the industry must have in order to do good science and persevere through all its disappointments.

However, in general the discourse surrounding the TechBio ‘frontier’ has been too nonspecific to be useful and too oversold versus what we actually have. From using ‘bio’ as a catch all for therapeutics, agriculture, consumer, infrastructure, and more, to even the premise of TechBio being ‘a direct application of engineering to biology’, it is difficult to understand precisely what anyone is saying these days or what separates the real advancements of today from the hopes and dreams of next century. It’s unsophisticated and conducive to hive-mind behavior. When no one knows what you are saying, it is impossible to be wrong. Biology is almost defined by its jungle of exceptions. The sooner we recognize how little we understand and have control over, the smarter we can be about putting the pieces of the puzzle together.

Hype is Just Hype

The fundamentals of TechBio rely on the convergence of advances in 3 areas: machine learning, synthetic biology, and high throughput automation. The thesis is that the combination of these advances and the network effects provided from an internal data moat can be strung together in a platform that continuously produces drugs. For discovery outfits like Recursion, the idea was that the platform provides materially higher success rates than industry by building a ‘map of biology’.

Reality has crushed this thesis. For each disease, new models are required, and little of our understanding scales from model to model. The activity of elements in biology is always context dependent, and it is currently impossible to recreate the entire context of human biology let alone between preclinical models of even related indications. Just to get cells to behave and grow in a dish requires them to be exposed to toxic levels of oxygen, glucose levels that are magnitudes higher than diabetic blood sugar, and uncharacterized reagents like fetal bovine serum that can vary from lot to lot and have no reason to be near the cell types they are cultured with physiologically. When cell lines are grown, they accumulate mutations and do not resemble the natural human networks and expression behavior of normal human cells. The reason why there is no ‘standardization’ of cell culture conditions is that any standard would be poor by default. Trying to industrialize this with technology, even just for a single model let alone many models in parallel, is a nightmare and unlikely to work. Finally, improved model systems such as iPSCs and organoids still often struggle to faithfully recapitulate underlying biology and are operator dependent in many cases. The starting cell population used to seed the system is a major contributor to experiment variation and the lack of reproducibility. Biology is not a playing field conducive to systematization.

Furthermore, even if the technological fundamentals of TechBio have progressed, the question of whether they are useful for solving problems in biology is always in question. Machine learning can be useful in certain spots such as imaging, SMILES, or sequence data but its utility past curve fitting and simple linear regression can be unclear for other applications simply because an algorithm cannot learn adequately when the underlying data does not represent the complexity of the problem. For example, our understanding of nucleic acid delivery is still quite poor and whether cell or organ specific delivery is entropy or ligand driven or mediated by other processes is not well understood. Attempting to use machine learning to solve this issue based on the chemical structure of the delivery vehicle thus has questionable value and is unlikely to offer an order of magnitude benefit over existing technology. There are too many disparate components of variation and acquiring all of them is too expensive and infeasible.

Synthetic biology, particularly CRISPR and being able to perturb biological systems is commoditized and can be better done in academia. The Broad Institute, Stanford University, and the NIH among other leading academic centers have constructed publicly available datasets that rival any of those that can be collected in industry. Across CMap, DepMap, ENCODE, TCGA, and many others is a wealth of expression data that provides at least 80% of whatever competitive advantage a next generation synbio screening startup may provide, while the remaining 20% of differentiation continues to be eroded by new projects. Academic centers have cheaper pricing, access to enthusiastic students, and the flexibility to collect data that may not be directly useful to company objectives.

High throughput automation technologies like plate handling and sequencing are useful only when asking relevant biological questions. Here again, screening facilities built into CROs or even organizations like Broad PRISM and NCI-60 already provide enough capacity to ask whether there is a need for TechBio companies to double check their work. Again, so much of problem solving in biology requires custom experiments to get useful data because the dimensionality is so high. CROs are surprisingly cheap and if a team is truly smart about designing experiments it’s unlikely that the type of speedup that high throughput will solve for is a limiting factor. HTS and sequencing progress may prove useful for discovery work, but are there examples of where this is better done in TechBio startups versus academia or large pharmas?

Academia and Big Pharma are Formidable Competitors

The next generation of PIs in academia and directors in pharma are frequently already up to speed on how to utilize machine learning, bioinformatics, lab automation, etc to modernize their experiments. At the very least, they are willing to try, and academic researchers have a multitude of advantages including scientific freedom, talented cheap labor, and access to collaborators and CROs at discounted rates to execute. The rise of alternative funding sources like Fast Grants, decentralized science grant distribution, and non-profit grants provide additional bandwidth to academic groups who wish to implement new technologies.

Overall, academia and big pharma aren’t poor and will continue to get more efficient as the biology stack improves. Big science projects that were once thought to only be possible with company funding are increasingly done by research centers and consortia like the NIH, Broad, and cross-institute collaborations. Importantly, the first to get a hold of these datasets are the academic centers that create them. They will be first to publish, patent, and start drug development. They will be most familiar with releases of new data, the quality with which data was generated, and have the ability to influence what additional projects will be initiated.

Large pharmaceutical companies in essence function as late stage VCs, hand picking technologies that synergize with existing internal capabilities. As a startup, you are not just competing against big pharma, you are competing against every other startup that pharma has interest partnering with. Big pharma also competes for the smartest talent, providing them higher salaries, access to the most up to date resources, and the opportunity to see their drugs move through the clinic if successful. Pharma scientists are quite smart, nimble, and capable. Large pharmas have the budgets to go after high risk targets and answer questions that no one else can, by directly funding their own clinical trials.

The connectivity of academic and pharmaceutical networks has gotten far stronger over the past few years and the pace of new ideas and findings makes it difficult to start a company off a single piece of technology. Being cognizant of who else is working on similar technologies both in academia and industry will become increasingly important; they are legitimate competitors. Claiming speedups in a startup environment over these players will get harder and harder to justify. Critically, very few startups who have claimed to be able to move faster than their peers in pharma actually have.

Why Haven’t You Published

TechBio isn’t a novel concept anymore. Given the promises of TechBio being able to accelerate discovery, allow us to learn more about biology than ever before, and discover patterns that ordinary scientists could not, we should expect TechBio companies to continuously generate and communicate new discoveries. Yet, these businesses rarely publish.

Notwithstanding publishing being a good practice of any citizen in science, the ability to put in writing and undergo peer review how a project was done and what the concrete outcomes and objectives were, is a major component of translational research. Participation in conferences, seminars, and other mechanisms of disseminating knowledge and data are essential activities for biotechs of all sizes. At the earliest stages, this may come via academic collaborators and at later stages the expectation is that the business should be generating its own data and communicating important results and discoveries.

The historical lack of involvement from TechBios with the academic community is concerning for several reasons. First, it allows for data driven inflection points and risk to be continuously pushed on to later investors. When poor data is unveiled, house of cards can quickly collapse (see Zymergen), wasting hundreds of millions of invested capital. Second, it removes scientists from any semblance of rigor or accountability when conducting experiments, potentially fueling a reproducibility and utility problem. Operator dependent results and lack of head to head comparison studies wastes capital. Finally, publishing data is simply good for the scientific community. It levels the playing field, ensures what the business is outputting is useful, and fulfills a core responsibility to the scientific community by helping train new researchers and disseminate findings. At the very least, we should expect some contribution to the pool of scientific knowledge if the pitch was good enough to merit millions in platform development funding.

Building a business to address any problem in biotechnology is difficult and as a community, we should recognize this. Foundational IP and publications should be a requirement for starting a new venture; a necessary bar for entry to prevent a flood of uninformed and non-differentiated science. The lack of published data even at the proof of concept level in many recently seeded TechBio companies is a red flag signaling a lack of patience, inability to validate, and poor focus. Publishing data does not deplete one’s first mover advantage if everything is covered by intellectual property, even Google and Facebook publish at rapid pace and in a field even more hyper competitive than biotechnology.

The Economics are Uninspiring

While private investors see TechBio businesses as platforms and thus insulated from risk associated with single asset therapeutics businesses, public markets disagree. Single asset or biology focused therapeutics are risky from the angle that if the hypothesis is wrong, the project gets shut down. Fortunately, capital requirements for these businesses are back loaded and trials are only greenlighted if the risk reward profile is favorable. In contrast, platform TechBio business have both front loaded and back loaded capital requirements. The costs of building, maintaining, and growing platforms can be enormous given that entirely new datasets need to be generated, labs built, and specialized talent hired. Recursion raised nearly $1 billion USD before ever conducting a clinical trial. Single asset businesses require investment for trials but are able to skip spending on discovery costs.

Biological risk is replaced by technical, platform risk which can be equally hard to mitigate. Platforms are fragile. They can lose utility, quickly getting replaced with the shiny next best thing just like any other asset. Before Illumina sequencing, Sanger sequencing was king, and after Illumina, will it be Nanopore or any of the other dozen Illumina killers that eventually topple the platform? Before, mRNA vaccines, DNA vaccines were mainstream, and maybe after a couple years will circular RNA or other types of RNA begin to replace mRNA? Zinc nucleases before CRISPR, and retrons after. Yeast display, then machine learning directed evolution, then phage assisted continuous evolution, and then what next? Platforms require longevity to pay off the costs of set up. If the pace of progress in biotechnology really is as fast as everyone says it is, platform moats will quickly disappear, as quick or faster than they materialized.

A common argument in favor of platforms is the idea that flywheel and network effects create a moat that allows TechBio companies to continuously develop new drugs, thus justifying platform costs. There is no proof of this. Even if it were true, it ignores the reason why we want drug companies to exist in the first place. The goal is not to approach drug development with a hammer looking for nails. A platform technology may only be (and often is) optimal for just a single indication and drug developers should be respectful of this. To remove biological risk, platform companies often opt for entering crowded indications with a multitude of existing approved or late stage assets to compete with, resulting in a bloated set of undifferentiated assets that can’t all be deployed. New drugs have known toxicity and unknown benefits and it is a net harm to society to develop them if we do not have high confidence that they will be practice changing. Unlike other industries, there are limits on consumption in healthcare; there is no hockey stick growth. Richard Murphey from Bay Bridge Bio insightfully points out that,

“[The] quality of R&D programs is more important than quantity. Most hits that come out of a screening platform are NPV negative. If quality is low, then the value of a platform decreases with scale. A valuable platform is one that can find a small number of valuable programs, not a large number of mediocre ones.”

The sad promise of platform TechBio therapeutics businesses is that the platform is worth anything meaningful at all. In any platform, value is almost entirely concentrated on the lead assets. There is the argument that TechBio platforms are derisked due to multiple streams of revenue from upfronts, milestones, and royalties. This can be true for services providers comfortable with reasonable capital outcomes, but as a discovery unit aiming for homeruns, your paycheck is directly proportional to whether you can develop drugs, no way around that. Signing away the economics of your most promising assets and the killer applications of the platform is no way to justify the expense of building the platform.

Any Thesis is About Picking Winners

The reality is that great companies are built in a multitude of different ways and there is no use forcing excess technology onto problems if it does not require it. A curious phenomena is that the further removed one gets from the earliest rounds of investment, the quieter the table pounding for TechBio becomes. Once you start thinking about valuations, milestones, and concrete business outputs, the ‘promise, potential, and power’ of the platform takes a back seat to things that clinicians and patients actually care about. Let’s retire the monikers and be more specific and precise when talking about the businesses we are building and why they are exciting.

Achieving a reasonable valuation and growth trajectory is about ignoring frenzied expectations and not being afraid to hit singles and doubles. Markets are more segmented and always smaller than expected with greater than expected competition. It pays to be leaner, take your time in a sandbox environment like academia, and be focused and honest about the limits and benefits of a technology and the rationale in turning it into a company. These must be businesses one day; at some point, data needs to translate to free cash flows.

It is unfortunate that founders can be compelled to shift out of a product focused mindset in order to raise money from platform obsessed funding sources. The point of building a platform is to graduate into a products business and if you can graduate early, you should! There is a notion that every company has to have a shot at being the next Genentech, a generational enduring company that continuously churns out blockbuster products. In reality, I think the point is to deploy capital in the most efficient way to help patients. It should be the expectation to be nimble and pick the low hanging fruit that academia has provided. Continuously starting, building, winding down once patients see benefit, not spending a penny more on wasteful development, and recycling the capital back into the ecosystem is a terrific result. Overall, there is no right way to do drug development and while TechBio shouldn’t be dismissed outright, it certainly isn’t a panacea. Any thesis is about picking winners; TechBio and Biotech are one in the same.

Why you should be a little bit stupid and a little bit lazy

2020-04-09T00:00:00+00:00

When I was at Johns Hopkins, Steve Wilson, one of my favorite mathematics professors, liked to say “if you want the best solution to a tough math problem, give it to your laziest friend”. He would often follow up and say “you know, the best mathematicians are a little bit stupid and a little bit lazy”.

As a student, I’d often chuckle hearing this. My high school and college friends often described me as being a little bit lazy. I wanted to maximize how much I learned, but realized learning didn’t perfectly correlate to performance in courses. As a result, my secondary goal was to hack the system to get good grades with minimal effort.

I hoped therefore that being a little bit lazy would help me be a great mathematician, or at least pass Wilson’s courses. It was only until recently that I fully grasped the truth of his statement. Being a little bit stupid and a bit lazy is not just a characteristic of good mathematicians, but a strategy that can be applied to all creative pursuits.

Being a bit stupid

If you’re super smart, you understand things as you learn them. You become deeply familiar with the research in your area. You quickly progress up the field. You learn all the and tools tricks.

Being a bit stupid has it advantages. You don’t immediately understand the fundamentals, so you question them in order to learn. You don’t read every paper in the field, so you aren’t as tied to preconceived ideas. Perhaps you aren’t as respected early on. You don’t write off ideas that seem crazy. You have freedom from high expectations, which gives you more mental bandwidth.

Many of the most important scientific discoveries have come from challenging previously held beliefs. Albert Einstein, who was never considered an exemplary student in his schooling, questioned Newtonian mechanics early on. He felt Newtonian mechanics could not reconcile with Maxwell’s equations for electromagnetism. This led to the development of his special theory of relativity, and later on, his general theory of relativity. Notably, Einstein developed the special theory of relativity while working in a patent office, as universities didn’t see him fit for an academic position. Questioning key assumptions allows the slightly stupid to uncover the unknown unknowns.

Being a bit lazy

If you’re hard working, you’ll attempt to use the vast machinery at your disposal. You immediately get to work on the problem. You work on it everyday till you solve it, which is admirable. A quality of great artists is their habit in always showing up to the easel everyday. When you do find a solution, it may be convoluted if you miss the essence of the problem.

Here’s what happens when you’re a bit lazy. You procrastinate, leading to greater creativity and original thinking as you subconsciously process. You seek the simplest solutions, which end up being more elegant. As a byproduct, your solutions are easier to explain. You boil down a problem to first principles, and avoid the trap of comparing to what exists. You solve your problem, and you have time for leisure.

Italian polymath Leonardo Da Vinci is a good example of a famous procrastinator. He is said to have had a horrible attention span, which pushed him to think and wander across fields. These distractions meant Da Vinci wouldn’t always finish his projects, but attained unique ideas balancing across science, math, and art.

The hardworking person usually solves the problem before the slightly lazy, and moves on to the next. It may seem she is solving more problems per unit of time, but this notion is only illusory. The slightly lazy will produce a cleaner and better solution.

Perhaps you have the frequent urge to procrastinate on your work by exploring the obscure. Maybe you don’t get always get concepts on the first try. If you often feel a little bit lazy and a little bit stupid, don’t look down on yourself, you’re in good company! If your behaviors are not destructive, embrace them instead - being a bit lazy and a bit stupid could be your biggest advantage.

Reflections on rapid learning of Spanish

2017-11-24T00:00:00+00:00

Two years ago, I spent my winter break living in Kibbutz Neot Semedar, a kibbutz in the Negev of Israel. One day I was tasked to design the electrical circuits with Dan, an Israeli-Argentinian 22 year engineer who recently finished his military service. One that day, something clicked. I took Spanish courses years ago, but always treated it as a class rather than a language. I never understood or could speak a thing. Dan had started teaching me Spanish from the perspective of a Hebrew speaker. He asked why I had an American accent in my Spanish given that I was from Israel. In that one day, I learned more Spanish than I did in 4 years of courses, and began to develop a theory on how one ought to learn a language. These last two months I’ve been living in Argentina, and successfully tested this theory.

This post is not about “why” to learn another language, but rather my “how”.

To preface, there are of course different levels of fluency. As it turns out, every Spanish speaking country has their own dialect. I speak Argentinian Spanish, due to my travels in Argentina and enamoration with Argentinian culture. I still have a ton to learn, and sometimes words just don’t come out for me. Regardless, to provide perspective on where my level of fluency is:

Most who come from outside of Latin America tend to perceive me as a local Argentinian.
Argentinians tend to think I’m European, usually someone from France/Brazil who learned Spanish with ease, or someone from another part of Europe who has Argentinian parents. I’ve met several English teachers in Argentina, who all found it easier to communicate with me in Spanish than in English.
The other day, I met two Chileans and hung out with them for a while. An hour in to our encounter, an Australian came and asked us a few questions in English. I answered, and was then complimented by two Chileans on how well I speak English. I tried to convince them after that I was in fact, not from Buenos Aires, but this was to no avail.

Obviously, there’s room where I can grow. Nonetheless, I think and speak in Spanish effortlessly. What I hope to distill is how I’ve been approaching new languages.

The theory for everyone

Try to remember when you first learned your first language. Do you find it hard to remember exactly how you initially learned? Very well. Try to remember how you learned in grade school. Perhaps the first time you learned that “will not” isn’t abbreviated as “willn’t” but rather as “won’t”? Or even in middle school, where you learned to use more advanced words, craft paragraphs, and the like.

Now, in these cases in your youth, do you ever remember consciously translating into another language or symbols of sort? Of course not. What you learned was just building off what you knew before, which you learned subconsciously by being listening and seeing. For verb conjugation, I’d be willing to guess you subconsciously learned the pattern and didn’t think to yourself “oh, I’m talking in the past tense and hence will now use the impreterite”.

Or even currently, if you encounter an unfamiliar word, do you translate it? I’d be willing to guess you either learn by carefully reading into its context or by using a dictionary.

When I met Dan, the Israeli-Argentinian backpacker, I realized I always sought to translate Spanish to English. It took someone translating Spanish into Hebrew to realize how wrong this is. Spanish isn’t English, nor is it Hebrew. Spanish is Spanish. Whatever language you want to learn is its proper self, not the language you already know. And as such, when you start to translate is when things start to go wrong.

This is a key point. You should avoid translating at all costs. Instead, seek to build complexity off of what you already know - in this fashion you will both learn new things and reinforce what you already know most naturally. You should seek to find yourself thinking unconsciously in your target language.

Assuming you’re reading this, you likely speak English at a very high level. Congratulate yourself! English is one of the most spoken and most useful languages in the world. Furthermore, you also speak a language of Anglo-Saxon and Latin roots, which makes it particularly easy to learn European languages. This is a huge advantage - you didn’t have this as a baby! For instance, you can instantly know that “resentimiento” in Spanish means resentment, without needing to go through the years of grade school Spanish education. This is your advantage to learning even faster.

Now, you did a few advantages as a baby. You weren’t really thinking at all in words and had mental silence. Research shows the incredible neural plasticity of babies. You had in a sense infinite time due to a lack of responsibility and more patience. And more than anything you had the necessity and desire to communicate.

While you can’t unfortunately regain the neural plasticity you had when you were a baby, you can simulate everything else. Learn to listen or shut off the voice in your head. Once you do this, you can start really listening to the new language. Then you can start thinking and speaking. This can also be trained by meditation.

Learn to have patience with your studies. Easier said than done, but try to view language learning as something without a deadline. Do a trip if you can, but if you can’t, find fun native speakers.

In the anthropological community, it’s common for anthropologists to go to a new area with an unwritten tribal language and become fluent in it. If you move somewhere where you don’t speak the language and have an intense interest in communicating with others, your chance of success will increase substantially. Stephen Krashen, professor emeritus at University of Southern California, refers to this as “comprehensible input”. I went to a talk of his at LangFest 2018 and loved it - leaving a short summary of his take here.

So in summary.

Build a base as fast as possible. This can be done fairly well with children’s movies - think Dora the Explorer or what you watched as a kid, but make it as compelling and interesting as possible. These shows are especially good, as they’re inherently immersive and take a long time. Duolingo, classes, can also be effective in building a base as long as you aren’t treating it like a class.
Afterwards, don’t translate.
always be thinking about the language if you’re thinking at all
Ask questions. People will notice if you really want to learn to interact with them and will gladly teach you.
Reread things you’ve read before in a new language.
Be super interested in the culture. This can’t be overstated. Have a necessity to speak. On this note, listen to music in the language. You’ll get the accent better, etc. Spanish has a ton of different music, from cumbia to reggaeton to bachata to salsa.
Accept corrections, if you don’t get corrected you don’t improve :)

Additions for bilingual speakers

I grew up in a Hebrew speaking household in the US. As such, I’ve spoken both languages fluently since I was young. You may have something similar if you’re reading this section - perhaps one parent is from a different country, perhaps you took French courses since your youth, etc.

What you don’t know is that as a result, you have an increased facility for language learning. This isn’t just because “you’ve done it before”.

When I first started learning Spanish, I chose to initially think of it as a mix of Hebrew (grammar, accent) and English (latin roots). I thought this approach could be sensible due to the map below - Spain is sort of in between Israel and England, where Hebrew and English emerged.

There’s perhaps some more logic to this, as Arabic influenced both Hebrew (due to proximity) and Spanish (due to the Moorish period). If I absolutely had to translate a word, I’d make sure to always translate into Hebrew, my less dominant language. I figured eventually I’d mentally separate Spanish from the two in my mind.

Although English + Hebrew doesn’t allow you to instantly pick up Spanish the way Hindi speakers may pick up Urdu, knowing both gives you an improved base to build off of. Take for instance the word “ya” in Spanish. In English, the concept of “ya” is a bit complicated to explain. Take these three sentences as an example:

Ya fui al supermercado - I already went to the supermarket.
Ya no sé qué hacer conmigo - I don’t know what to do with myself anymore. (This is also the name great song by Uruguayan band Cuarteto de nos)
Ya vuelvo - I return soon.

Translating these directly into English causes immediate problems. In the first sentence, “ya” means “already”; in the second, “ya” means “anymore”; in the third, “ya” means “soon”. In Hebrew, on the contrary, “ya” is exactly the word “כבר” in these common contexts. Or take the difference between “conocer” and “saber” which is confusing to explain in English but instantly translatable in Hebrew to “להכיר” and “לדעת”. It would seem plausible this triangulation to a new language applies beyond just Spanish or even just Romance/Germanic Language.

What I did + resources + suggestions

Watch Naruto with Spanish subtitles/dubs when available. If I didn’t understand a word I’d keep going in effort to subconsciously pick up new words - I already knew the context and general theme of Naruto.
Language exchange with a great friend and polyglot who was looking to learn Hebrew and already spoke fluent Spanish.
Listen to a lot of music in Spanish. Listening attentively and singing along slowly improved my pronunciation and accent.
Try to talk to everyone in Argentina (works best in non-touristy areas). To me, I was (and still am) with the culture and language in Argentina. Hence, every opportunity speaking to someone was an opportunity to learn more. People at stores. Bus drivers. Store owners. If you really make an effort to try to know them and speak their language, they notice.
Couchsurfing! With Couchsurfing I stayed in the homes of 10-20 people, each of whom not only showed me their culture but also were willing to correct me as I spoke to them in their native tongue.

Ending note

Learning a new language has completely changed my life. Beyond simply the intellectual challenge, it opens a pathway to make millions of new friends from all over the world. My life would not be the same without the wonderful friends I have made in Argentina. To anyone learning - I hope my take helps, stick through the process as it becomes incredibly rewarding in the end.

Why MedHacks?

2015-09-28T00:00:00+00:00

MedHacks, the medical themed hackathon I started at Johns Hopkins, is next week! For details on the motivation behind MedHacks, check out our most recent Medium post.