Sample Recommendation Form
STEM Enhancement in Earth Science (SEES) Summer Intern Program
Cockrell School of Engineering, University of Texas at Austin
Program specifies: Students should possess appropriate reading, writing, math, computation, and scientific analysis skills. We wish to provide this opportunity to students who may not otherwise have the opportunity to participate in a demanding but rewarding STEM opportunity.
- How long have you known the applicant and in what capacity?
I teach Computer Vision, an advanced Computer Science course at Thomas Jefferson. Charlie is a student in my class this year, I have known him through our regular interactions.
- What three words come to mind as you consider the applicant? Why?
Rigor: Not only has Charlie maintained a perfect GPA, he chooses the most challenging courses available, years ahead of traditional curriculum. Having achieved USACO platinum, I have no doubt his programming skills have enabled him to pursue advanced research projects.
Integrity: Charlie has a strong work ethics and scientific integrity. He wasn’t satisfied with running the code but also tries to understand the deeper meaning of how it connects to the real world. For instance, during ASSIP internship, he uncovered that the activation of latent features really encoded for tumor cells. In his RNA drug screening research, he incorporated confidence level indicators to let the model reveal when it is uncertain and improve practical use.
Creative: He likes to draw connections across topics. With the CS concepts he’s learned in my class and in his independent studies, he connected that with concepts in molecular biology to find solutions in the RNA binding problem. His RNA/Molecule binding affinity screen model has several inventive solutions, such as replacing the need for 3D structures using evolutionary data and LLM embeddings, or adapting mechanisms from vision-language models.
- What traits does this student possess that would make him/her a qualified SEES intern?
Courage in science: His research moved focus from mainstream protein-targeting to an offbeat path of RNA-targeting, a less trodden area which might make it appear formidible, but is actually more fertile for improvement. If successful, I have no doubt about the impacts it could have on drug discovery.
Grit: Charlie put in consistent self-directed study and discipline into his pursuits of his extracurricular studies like USACO, AIME, as well as his research projects. Implementing training scripts for the RNA model, troubleshooting, and waiting hours for one run results to show was definitely a long and arduous process full of obstacles.
Compassion: He contributed to collaborative learning in class through active participations, helping out peers with questions and having lively discussions.
- Please provide any additional information about this student that will help in our selection process.
Through our interactions, Charlie comes across as a student with technical intensity and a genuine passion for solving real world problems on a larger scale. His time management and methodical mindset will very much let him acquire new knowledge and skills. He is also very active in TJ’s math, CS, and physics clubs, that likely inspired much of his interest in the interdisciplinary research he conducts.
Charlie’s focus is not limited to STEM. Besides insights from books in aging and longevity, he’s also a strong writer. His essays received consecutive awards in the Scholastic Writing Contest and was recognized as semifinalist in the John Locke Essay Competition.
I recommend Charlie without reservation for the NASA SEES program. He has great quantitative skills, research abilities, and maturity in collaboration that will let him thrive at SEES, and I’m sure he will contribute meaningfully to your community.
Process of the research
(very OPTIONAL. Just wrote this here for your reference, in case you were wondering)
I. Beginning
I have already visited the emergency room more than 20 times due to a rare skin condition called Ehler’s Danlos Syndrome (hyperfragile skin). Even on a simple hiking trip, all it took was one sharp twig and all my parents’ hard-earned savings wasted on stitches and late nights in the ER.
The burden had become too much, so I took to Google Search to find a treatment. Initially, all I found were still symptom-level treatments, like braces or creams, not therapies. However, from reading deeper, I discovered RNA drugs as a promising research direction. Modern drugs target proteins, but long, fibrous collagen proteins of the skin can’t be drugged, whereas the upstream col5a1 RNA can.
Yet, I was surprised to find RNA drug discovery still in its infancy, where we need to develop accurate computational prescreening before physical experiments become tractable cost-wise.
II. Inspirations
For context, my childhood hero is Demis Hassabis, who led Google DeepMind in creating AlphaFold (protein structure prediction model) — so much so that my favorite movie as a kid was the AlphaGo documentary. His audacity to strive directly towards solving all disease and uncovering how reality works was an alluring attitude and is what impelled me to learn ML. I remember thinking: I want to be next Hassabis. (still do, actually).
More recently, I also discovered Laura Deming and her work in longevity. She’s introduced to me so many wonderful resources for learning biology, particularly book.bionumbers.org and Molecular Biology of a Cell, which I devoured like a little kid with chocolate cake. I thought, finally someone who understands the rapturous excitement of intuiting science.
Wanting to follow Hassabis’s footsteps, I spent all my time outside school (and sometimes also in school) over the years first learning calculus, then the ML basics, then teaching the computer to play soccer, or recognize my face, among various other little projects. With each step I grew more and more curious and fascinated.
There’s no feeling quite like marvelling as the computer begins to learn the way we do. It’s just matrix multiplications! And then I sit on a cushion and watch my own thoughts arise, and plant avocados in my yard and watch a tree sprout out of nothing. Struggling to quantify how life and thought arise (also just multiplications?), the mystery gnaws at me.
To sum, my lifetime obsession has just been dying to uncover the source behind this wonderful thing called experience and help us live healthier to improve it.
III. From learning to innovation
As I first dug into the latest computational models for RNA drug screening, I was reminded of AlphaFold and became super hooked that very weekend.
I thought this would be a really great learning exercise — to replicate impactful papers and practice PyTorch in a field so close to my life (EDS and AlphaFold). So I organized this into my active project. Through evening hours and all of summer, I played around with different model architectures.
I started with basic MLP and gradually added features and tested them on Colab to get a feel for what worked and why. e.g., RNA-FM, RiNALMo, Graphormer, UniMol as sources of pretrained embeddings; secondary structure and GNNs; early vs late fusion; and so on. In a few months, I finished trying out much of the fundamental techniques seen in existing binding affinity models, so I started looking to adjacent fields.
That’s when I went back to AlphaFold and remembered its evolutionary MSA module (called Evoformer). Evoformer is largely what made their structure prediction so successful. But looking at the top models like DeepRSMA, I thought “where’s the MSA module?” Thinking it would be good practice to code up something without a clear answer, I decided to just give it a shot.
… and the model got better!
That moment changed everything! I realized improvements really were possible. So since then, in the last seven months, I became deeply deeply obsessed with this project and the idea that it might really turn into something useful. Late evenings, weekends, and bus rides to school blurred by, as I dug for insights in papers from both near and far fields, and tried adapting them towards the RNA ligand problem. The project would go on to span almost 1000 hours.
IV. Challenges
The main difficulty was developing an accurate model despite scarce 3D RNA structures for it to learn from.
My first instinct was to just find more data somewhere. I spent weeks scraping databases, trying to combine different sources. But the quality was terrible — inconsistent labels, different experimental conditions. It made the model worse, not better.
I considered using predicted 3D structures but the models were far to bulky to fit into my pipeline and iterate on designs. And their often inaccurate results propagate, ruining my final prediction.
Finally, I stepped back to consider first principles: there’s no physical law that required explicit structures in binding affinity prediction (where we just model how well RNA and drug bind, not the exact pocket, which is enough for screening drug candidates). Could we surrogate 3D structural context by effectively fusing rich alternative sources?
I spent weeks trying different architectures — prenorm versus postnorm, positional encodings, early fusion with graph neural networks. Each training run took a whole day on Colab’s strong A100 GPUs (fortunately free for students). And more than once, I’d wait 24 hours only to discover a bug at the very end.
Finally, I took inspiration from the actual binding process (where each atom on the RNA and molecule simultaneously influence each other) and designed a symmetric co-attention with tanh gating that was partly inspired by DeepMind’s Flamingo vision model. Combined with my tri-modal embeddings and a new uncertainty-based semi-supervision, the model finally broke through.
Another important challenge was making sure it was learning the right thing. At first I used train/test splits by random and iterated model designs using those benchmarks as the metric. But I later realized it was just memorizing answers this way, so I started splitting by molecule, RNA, and disease types. Performance initilaly fell, but the resulting model was able to demonstrate much better diversity to a broader range of applications. Excellent! Science at work.