Videos -561cZmir5Q

From Tokens to Cells: Foundation Models for Single-Cell Biology - Akram Baharlouei, Altos Labs

index_state ready data_status ok

AI Engineer· published 2026-07-19· 0:16:57· en-US· indexed 2026-08-10 19:42

Open on YouTube

43 shot(s).

keyframes kept every frame deduplicated

What was stored

cues: 142; whisperx 142
chunks: 29; from 142 cues
keyframes: 30; kept of 43 captured
frames with text: 30; 725 lines read
chapters: 7; from the source metadata
keyframe bytes: 3.7 MB; word timings on 142 cues

Provenance

Each pipeline stage, its state and the model that produced it
stage	state	model	started	took
`fetch`	done	—	2026-08-09 23:38	0s
`stt`	done	—	2026-08-09 14:52	19s
`chunk`	done	—	2026-08-09 14:53	0s
`text_embed`	done	—	2026-08-10 19:42	0s
`keyframe`	done	—	2026-08-09 14:53	1m 03s
`ocr`	done	—	2026-08-09 14:54	18s
`frame_embed`	done	—	2026-08-10 19:42	5s

Frames, and what the machine read

0:06 #0 done14 line(s)

shot 0·sharpness 741.4
1. localhost:8765/From Tokens1.00
2. X0.87
3. +0.97
4. Ask Gemini1.00
5. ←0.93
6. C0.57
7. localhost:8765/From%20Tokens%20to%20Cells%20v4.dc.html#11.00
8. AI ENGINEER TALK1.00
9. From Tokens to Cells0.99
10. The engineering challenges of foundation models for single-cell biology1.00
11. Akram Baharlouei1.00
12. Altos Labs1.00
13. 1/281.00
14. Reset1.00
0:33 #1 done23 line(s)

shot 1·sharpness 691.4
1. S0.63
2. localhost:8765/From Tokens1.00
3. +0.95
4. Ask Gemini1.00
5. ←0.98
6. localhost:8765/From%20Tokens%20to%20Cells%20v4.dc.html#21.00
7. OUTLINE1.00
8. akram baharloue1.00
9. 011.00
10. Why single cells1.00
11. What to measure0.99
12. 020.88
13. Why it's hard: heterogeneity1.00
14. 030.98
15. Modeling cells1.00
16. 040.99
17. Takeaways1.00
18. 051.00
19. FROM TOKENS TO CELLS1.00
20. ALTOS LABS1.00
21. 2/280.97
22. Reset1.00
23. R1.00
0:39 #2 skipped

shot 2·duplicate of #0
1:01 #3 skipped

shot 3·duplicate of #1
1:07 #4 done13 line(s)

shot 4·sharpness 602.8
1. localhost:8765/From Tokens1.00
2. X0.82
3. +0.97
4. Ask Gemini1.00
5. ←0.92
6. localhost:8765/From%20Tokens%20to%20Cells%20v4.dc.html#31.00
7. SECTION 011.00
8. Why single cells0.99
9. Rejuvenation, the path to a virtual human, and a way to bend the cost0.99
10. curve of medicine.1.00
11. 40.70
12. 3/281.00
13. Reset1.00
1:17 #5 done24 line(s)

shot 5·sharpness 688.8
1. localhost:8765/From Tokens1.00
2. X0.75
3. +0.96
4. Ask Gemini1.00
5. K0.73
6. C0.89
7. G0.69
8. localhost:8765/From%20Tokens%20to%20Cells%20v4.dc.html#41.00
9. CELLULAR REPROGRAMMING1.00
10. Yamanaka factors reset a cell to a stem-like state1.00
11. OSKM0.97
12. ALTOS1.00
13. OSKM binding DNA — reprogramming0.97
14. YOUNG1.00
15. PAUSE1.00
16. REPLAY1.00
17. REC1.00
18. Oct40.99
19. Sox21.00
20. ●Klf40.83
21. c-Myc0.99
22. Oct4 - Sox2 - Klf4 - c-Myc are transcription-factor proteins. They enter an aged cell,0.97
23. its epigenetic age toward an induced pluripotent stem cell.0.99
24. Auto-orbit; click the scene to play / pause.0.99
1:59 #6 done21 line(s)

shot 6·sharpness 668.3
1. localhost:8765/From Tokens1.00
2. X0.84
3. +0.98
4. Ask Gemini1.00
5. K0.87
6. C0.77
7. G0.78
8. localhost:8765/From%20Tokens%20to%20Cells%20v4.dc.html#41.00
9. CELLULAR REPROGRAMMING1.00
10. Yamanaka factors reset a cell to a stem-like state0.99
11. OSKM0.98
12. ALTOS0.99
13. OSKM binding DNA-reprogramming0.98
14. YOUNG1.00
15. PAUSE1.00
16. REPLAY1.00
17. REC1.00
18. Oct41.00
19. Sox21.00
20. Oct4 - Sox2 - Klf4 - c-Myc are transcription-factor proteins. They enter an aged cell, dock onto DNA in the nucleus, and rewind its epigenetic age toward an induced pluripotent stem cell.0.99
21. Auto-orbit; click the scene to play / pause.0.97
2:17 #7 done25 line(s)

shot 7·sharpness 660.2
1. localhost:8765/From Tokens0.98
2. X0.76
3. +0.98
4. Ask Gemini1.00
5. ←0.87
6. C0.55
7. localhost:8765/From%20Tokens%20to%20Cells%20v4.dc.html#41.00
8. CELLULAR REPROGRAMMING1.00
9. Yamanaka factors reset a cell to a stem-like state0.99
10. OSKM0.98
11. ALTOS0.98
12. akram baharlouei0.99
13. Induced pluripotent stem cell(iPSC)0.97
14. ERIGENETICAGE0.94
15. OLD0.99
16. YOUNG1.00
17. PAUSE1.00
18. REPLAY1.00
19. REC1.00
20. Oct40.99
21. Sox21.00
22. Klf40.92
23. ●c-Myc0.91
24. Oct4 · Sox2 - Klf4 - c-Myc are transcription-factor proteins. They enter an aged cell, dock onto DNA in the nucleus, and rewind its epigenetic age toward an induced pluripotent stem cell.0.99
25. Auto-orbit; click the scene to play / pause.0.97
2:59 #8 skipped

shot 8·duplicate of #7
3:10 #9 done26 line(s)

shot 9·sharpness 532.6
1. localhost:8765/From Tokens0.99
2. X0.72
3. +0.97
4. Ask Gemini1.00
5. ←0.99
6. C0.54
7. localhost:8765/From%20Tokens%20to%20Cells%20v4.dc.html#51.00
8. 01. WHY SINGLE CELLS0.96
9. How four factors reprogram a cell — step by step0.99
10. STEP1.00
11. akram baharloue1.00
12. STEP 00.97
13. Old skin cell1.00
14. skin genes1.00
15. ON1.00
16. stem genes0.99
17. = OFF0.89
18. SKIN0.93
19. 40.65
20. old skin cell1.00
21. I1 PAUSE0.89
22. REPLAY1.00
23. REC1.00
24. old skin cell - iPSC . -2-4 weeks in vivo0.94
25. 5/281.00
26. Reset1.00
3:40 #10 done24 line(s)

shot 10·sharpness 843.2
1. localhost:8765/From Tokens1.00
2. +0.98
3. Ask Gemini1.00
4. ←0.61
5. localhost:8765/From%20Tokens%20to%20Cells%20v4.dc.html#61.00
6. 01· WHY SINGLE CELLS0.99
7. Build the virtual human like a pyramid1.00
8. akram baharloue0.99
9. FOUNDATION1.00
10. LAYER1.00
11. PEAK1.00
12. SCALE1.00
13. Cell1.00
14. Tissue0.99
15. Organ1.00
16. Body1.00
17. Genes, transcripts, proteins0.98
18. Cells talking to neighbors0.98
19. Tissues strung together1.00
20. A simulated population0.99
21. Master the simulation of one unit and you gain the code to simulate the organism. (Sarah Teichmann & the0.99
22. Human Cell Atlas.)1.00
23. FROM TOKENS TO CELLS1.00
24. ALTOS LABS0.99
4:13 #11 skipped

shot 11·duplicate of #10
4:18 #12 done25 line(s)

shot 12·sharpness 383.8
1. S0.69
2. localhost:8765/From Tokens1.00
3. X0.64
4. +0.95
5. Ask Gemini0.99
6. ←0.98
7. C0.67
8. localhost:8765/From%20Tokens%20to%20Cells%20v4.dc.html#70.99
9. THE SCALING STRATEGY1.00
10. From one cell to a virtual human1.00
11. MASTER ONE CELL COMPOSE TO ANY0.99
12. akram baharlouei1.00
13. Single cell1.00
14. 11 PAUSE0.89
15. REPLAY1.00
16. ● REC0.84
17. neuronblood1.00
18. immune0.98
19. epithelial1.00
20. muscle1.00
21. One cell = a gene expression vector in -20,000 dimensions.0.99
22. g organ -> full virtual human.0.97
23. Master one cell - compose to any scale.1.00
24. 7/281.00
25. Reset1.00
4:49 #13 done33 line(s)

shot 13·sharpness 1648.2
1. localhost:8765/From Tokens1.00
2. X0.51
3. +0.97
4. Ask Gemini1.00
5. C0.57
6. localhost:8765/From%20Tokens%20to%20Cells%20v4.dc.html#81.00
7. 01· WHY SINGLE CELLS0.98
8. Moore's Law vs Eroom's Law: the contradiction1.00
9. MOORE'S LAW ↑ Technology1.00
10. EROOM'S LAW ↓ Drug Output0.99
11. Computing power doubles every ~2 years1.00
12. Drug output halves every ~9 years since 19501.00
13. 10*× more0.91
14. 10× fewer0.96
15. TRON0.92
16. Transistors per chip→ exponential improvement0.98
17. 19700.99
18. 19821.00
19. 19941.00
20. 20061.00
21. 20161.00
22. 20241.00
23. New drugs per $1B R&D → exponential decline0.96
24. 19501.00
25. 19771.00
26. 20061.00
27. 20200.99
28. As computing improved exponentially, drug productivity got worse exponentially.0.99
29. Single-cell AI is the lever to bend this curve back.1.00
30. FROM TOKENS TO CELLS1.00
31. RESTART1.00
32. PAUSE1.00
33. ● REC0.85
5:03 #14 done56 line(s)

shot 14·sharpness 1113.6
1. S0.56
2. localhost:8765/From Tokens1.00
3. +0.97
4. Ask Gemini1.00
5. ←0.99
6. C0.73
7. localhost:8765/From%20Tokens%20to%20Cells%20v4.dc.html#91.00
8. Al at every stage of drug development0.99
9. akram baharlouel0.97
10. Research &1.00
11. Preclinical1.00
12. Clinical1.00
13. Review &0.95
14. Development1.00
15. Studies1.00
16. Trials1.00
17. Approval1.00
18. 3-6 years0.99
19. 1-2 years0.97
20. 4-7 years1.00
21. 1-2 years0.98
22. o0o0.54
23. AAA0.95
24. Target identification0.98
25. In vitro studies1.00
26. Phase I, II & III trials0.98
27. Safety & efficacy review1.00
28. Compound screening1.00
29. In vivo animal models1.00
30. Dosage & safety monitoring1.00
31. Regulatory approval1.00
32. Lead optimization1.00
33. Toxicity testing1.00
34. Patient recruitment0.99
35. Post-market surveillance1.00
36. ↓1.00
37. ↓0.99
38. ↓1.00
39. →0.99
40. AI REPLACES1.00
41. AI REPLACES1.00
42. AI REPLACES0.96
43. AI REPLACES1.00
44. Virtual cells1.00
45. Virtual humans1.00
46. Digital twins1.00
47. Protein engineering models1.00
48. Virtual tissues & organs1.00
49. In silico trials1.00
50. Adverse event prediction0.99
51. Scientific reasoning agents0.98
52. In silico toxicity0.98
53. Synthetic patient cohorts0.99
54. Continuous monitoring1.00
55. FROM TOKENS TO CELLS1.00
56. ALTOS LABS1.00
5:43 #15 skipped

shot 15·duplicate of #14
6:00 #16 done10 line(s)

shot 16·sharpness 520.1
1. localhost:8765/From Tokens0.98
2. X0.77
3. +0.99
4. Ask Gemini1.00
5. ←0.99
6. localhost:8765/From%20Tokens%20to%20Cells%20v4.dc.html#101.00
7. SECTION 021.00
8. What to measure0.99
9. Different ways to read a cell — and why we focus on one.0.99
10. Q0.70
6:05 #17 done46 line(s)

shot 17·sharpness 1454.7
1. localhost:8765/From Tokens1.00
2. +0.97
3. Ask Gemini1.00
4. ←0.58
5. G0.59
6. localhost:8765/From%20Tokens%20to%20Cells%20v4.dc.html#111.00
7. Many ways to read a cell1.00
8. One cell — many measurement modalities, each capturing different biology.0.99
9. akram baharlouei1.00
10. SCRNA-SEQ OUR FOCUS0.97
11. Morphology1.00
12. Transcriptome1.00
13. ~20,000-gene vector per cell - high throughput - millions of cells0.97
14. cell shape & structure0.99
15. Epigenomics1.00
16. X0.94
17. GENOMICS1.00
18. The DNA code0.99
19. chromatin access1.00
20. Static sequence · same in every cell- genome-wide variants0.97
21. Genomics1.00
22. EPIGENOMICS1.00
23. DNA sequence0.99
24. scRNA-seq *0.94
25. Which DNA regions are open - histone modifications0.99
26. Chromatin accessibility1.00
27. transcriptome1.00
28. PROTEOMICS1.00
29. Protein output0.99
30. Actual functional molecules· low throughput · expensive0.98
31. SPATIAL1.00
32. Location in tissue1.00
33. Proteomics1.00
34. Gene expression + tissue coordinates0.99
35. ell neighborhoods0.96
36. protein output1.00
37. MORPHOLOGY1.00
38. Cell shape & structure1.00
39. Spatial1.00
40. Size, shape, organelle layout - light & electron microscopy0.99
41. location in tissue0.99
42. IMAGING1.00
43. Fluorescence & microscopy1.00
44. Protein localization, live-cell imaging · high-content screening0.98
45. FROM TOKENS TO CELLS1.00
46. ALTOS LABS1.00
6:37 #18 skipped

shot 18·duplicate of #17
7:12 #19 skipped

shot 19·duplicate of #17
7:50 #20 skipped

shot 20·duplicate of #17
8:14 #21 skipped

shot 21·duplicate of #17
8:43 #22 done12 line(s)

shot 22·sharpness 689.1
1. 60.56
2. localhost:8765/From Tokens1.00
3. X0.80
4. +0.98
5. Ask Gemini1.00
6. ←0.97
7. G0.54
8. localhost:8765/From%20Tokens%20to%20Cells%20v4.dc.html#121.00
9. akram baharlouei1.00
10. SECTION 031.00
11. Why it's hard: heterogeneity1.00
12. Two “identical” cells rarely read the same — and not all of it is signal.0.98
9:32 #23 done34 line(s)

shot 23·sharpness 607.9
1. S0.64
2. localhost:8765/From Tokens1.00
3. X0.58
4. +0.97
5. Ask Gemini0.93
6. ←0.63
7. G0.51
8. localhost:8765/From%20Tokens%20to%20Cells%20v4.dc.html#131.00
9. WHY SINGLE-CELL DATA IS NOISY1.00
10. A cell is a movie — we only capture snapshots0.99
11. POPULATION . SCRNA-SEQ SNAPS0.95
12. akram baharlouei1.00
13. scRNA-seq captures a snapshot1.00
14. each cell frozen at a different moment in its cycle0.99
15. same tissue . different gene expression = heterogeneity0.97
16. G11.00
17. G11.00
18. G11.00
19. M1.00
20. S0.61
21. S0.80
22. G11.00
23. G11.00
24. M1.00
25. C0.98
26. G11.00
27. G11.00
28. G11.00
29. S0.82
30. I1I PAUSE0.84
31. REPLAY0.98
32. • REC0.81
33. A cell continuously cycles: grow → copy DNA → split chromosomes → divide. scRNA-seq destroys the cell and reads its mRNA at one frozen moment - giving thousands of disconnected0.99
34. snapshots at random cycle positions. That is the core source of noise and heterogeneity.0.99

Transcript

142 cues· 2,339 words· 12,729 chars

chunk 0 · 0:12–0:56 · 103 words · 579 chars
0:12 Okay, let's get started.
0:14 My name is Akram.
0:15 I'm a machine learning engineer at Altos Labs.
0:18 Altos Labs is a biotech startup and the goal is to restore cell health and resilience through cellular rejuvenation to inverse disease and disabilities that can happen throughout the life.
0:32 And the title of my talk is From Tokens to Cells.
0:37 And this is my kind of view as someone without bio background to kind of looking into the engineering challenges of foundation models for single cell biology.
chunk 1 · 0:51–1:35 · 110 words · 589 chars
0:51 And what I want to talk about first, what is single cell?
0:54 Why do we care about single cell?
0:56 How do we measure it?
0:57 What are the problems with the data, getting the data?
1:01 And then looking at current state of the art foundation models and then some takeaways in the end.
1:09 So what is single cell and why do we care about it?
1:12 I would like to start with my favorite example, Yamanaka factor.
1:17 In 2006, Shinya Yamanaka discovered four transcription factors.
chunk 2 · 1:22–1:51 · 61 words · 334 chars
1:22 There are four specific type of proteins that when they were overexpressed in a cell,
1:30 They did it in a cell, which was an age-old skin cell.
1:36 It could reprogram the skin cell, old skin cell, to embryonic stem cell-like state.
1:42 So basically, from cell type, skin cell, old, it could reprogram it back to a young embryonic stem cell type.
chunk 3 · 1:54–2:31 · 91 words · 510 chars
1:54 And this was a breakthrough for biology and then got him Nobel Prize later in 2012 because of the application and the new chapters possibilities for medicine, from regenerative medicine for like to be able to kind of regenerate tissues or organs when we can like reprogram a specific cell to any cell that we want to partially programming an aging application.
2:22 For partial programming specifically is that we can also turn out that we can also only change the age of the cell.
chunk 4 · 2:29–3:08 · 101 words · 504 chars
2:29 We don't have to change the type.
2:31 So by changing the age of the cell, the goal is, the hope is we can restore some of youthful function that we lose throughout life as we age.
2:41 And then kind of like having this as a medicine, maybe like an mRNA medicine.
2:47 for curing the disease that we encounter as we age.
2:51 And that's one example.
2:54 And then I think this year, 2026, we got the first type of reprogramming medicine.
chunk 5 · 3:00–3:42 · 102 words · 525 chars
3:00 I think it's OSK.
3:03 And it's going to be tested in a human.
3:06 So it took kind of like 20 years.
3:11 And the other reason is that why we want to study a single cell is that we would like to model to have this unified holistic view of single cell to be able to model cell.
3:25 And from there, hopefully we can model tissue
3:28 An organ and then the entire human.
chunk 6 · 3:31–4:08 · 79 words · 478 chars
3:31 And there are like projects like Human Cell Atlas that actually started this effort, this initiative.
3:37 And then they like, you know, they have mapped every single cell inside human.
3:44 And the goal ultimately is something like maybe for ultimately after some long time, we can actually model human body.
3:56 Or there are terminologies like virtual cell, virtual tissue, virtual human, virtual twins.
chunk 7 · 4:02–4:47 · 102 words · 552 chars
4:02 They all saying the same thing that the more that we can model these living organisms,
4:10 the better we are in understanding our body and then how we can treat medicine we can develop drugs and the other problem is that so we you know I think for this audience we know about Moore's law the compute is getting like double every year and then we have exact opposite on drug development
chunk 8 · 4:35–5:15 · 103 words · 575 chars
4:35 Basically, the number of drugs that develop each year is kind of declining, which is surprising with all the advances in technology, in AI.
4:45 And this is surprising to see.
4:47 And then drug development is a field that failure is, you know, very normal.
4:50 Maybe the acceptance rate is kind of like, you know, 5% or even less.
4:57 And when we're looking at drug development pipeline from the early research and development all the way to preclinical trial and then the final stage, the whole pipeline, it could take up to 10 years easily, and then it costs millions.
chunk 9 · 5:14–5:53 · 108 words · 577 chars
5:14 It can cost billions.
5:16 And the goal here is that with the advances of AI and also like with this virtual cell, virtual organ, virtual human, the goal is that we can kind of like reduce this time, improve, reduce this timeline and
5:32 and also we're looking at the entire timeline let's say if we only look at like research and development the models like you know proton design and stuff we might like save like you know few years here but at the end maybe it's not gonna help uh for the entire pipeline so it's important to have innovation and a breakthrough uh across all pipeline
chunk 10 · 5:56–6:40 · 107 words · 589 chars
5:56 Now that we know single cell is important, how can we measure single cell?
6:03 So looking at single, yeah, it's amazing.
6:05 This is just one single cell.
6:07 There is a lot going on inside that one little tiny organism.

Chapters

0:00 Introduction to single cell
1:09 Significance of single cells
5:56 Measuring biological data
8:29 Data heterogeneity challenges
9:51 Transformer-based models
13:42 Flow matching for cells
15:12 Conclusion and takeaways